Latest — 22 Oct 2026. Tokyo had its worst session of this watchlist’s year and closed on the low of the day. The Nikkei closed 62,907.84, down 1,713.40 points or 2.65%, confirmed against the 22 October daily bar, regularMarketPrice stamped 15:45 JST, and independently against Il Sole 24 Ore, which put Tokyo at the same 2.65%. The index traded as high as 64,866.44 in the morning, up 0.38% on the day, then settled 1,958 points lower at the session’s weakest print. Every Tokyo line on this page did the same thing and closed at or within a yen of its own low: Harmonic Drive −9.55% to ¥4,875 on 1.71 times average volume, Nabtesco −4.38% to ¥4,166 on 1.02 times, TDK −4.20% to ¥2,780.5 on 0.83, Yaskawa −4.07% to ¥4,033 on 1.08, Nidec −2.67% to ¥1,715 on 0.59 and Aichi Steel −0.76% to ¥2,857 on 0.89. Three are period lows: Yaskawa’s is its lowest close since 31 March 2026, Harmonic Drive’s since 28 April 2026, and Nidec’s ¥1,715 is the lowest close the name has printed in at least five years. No company on this page filed anything dated 22 October. Six days remain to the 28 October anniversary.
It was not a handoff from New York, which is the part worth recording. The 21 October US session rose on every index this page checks: the S&P 500 +0.58% to 7,757.17, the Nasdaq Composite +1.05%, the Dow +0.69% and the PHLX Semiconductor index +1.97%, with Nvidia itself up 2.54%. Tokyo opened higher into that and then fell 2.65%. The damage was concentrated in chip equipment and the AI complex rather than in automation as such: Tokyo Electron −6.99%, SoftBank Group −6.86%, Advantest −6.01% and Murata −4.80%, against Keyence −0.62% and Daikin −0.47%. The two automation bellwethers sat in between, Fanuc −3.42% and SMC −4.20%, which puts Harmonic Drive’s fall at 2.8 times Fanuc’s and 3.6 times the index. That is the ordinary behaviour of the highest-multiple name in a complex being de-rated rather than evidence of a disclosure, and the name’s own record says so: Harmonic Drive has now fallen nine and a half per cent or more on six sessions since 1 July, and this was the smallest of the six. The previous one was −11.64% on 15 October.
China put the yttrium controls on Japan into its own words, and the page can quote them rather than infer them. Asked at the 21 October Ministry of Foreign Affairs press conference about a Financial Times report that Japan is months from exhausting its yttrium, spokesperson Mao Ning answered in four sentences, of which the second is the one that matters: “China implements export controls on Japan in accordance with the laws and regulations. The aim is to thwart Japan’s remilitarization and attempt to possess nuclear weapons.” She added that China’s position on keeping critical-mineral supply chains “safe and stable has not changed”. That is taken from the ministry’s own published transcript, not from coverage of it. Around it: the FT reported on 20 October that Tokyo has asked the US government and American companies to press Beijing; Chinese customs data put yttrium oxide exports to Japan at 4,000 kg in September against 51,000 kg a year earlier; a Japanese industry ministry official said exports have stalled worldwide rather than only to Japan; and Argus has yttrium oxide at about $1,150/kg outside China against roughly $9/kg inside it. Two qualifications keep this honest for this page. Yttrium is not an NdFeB input: it goes into jet engines and chipmaking equipment, so it does not touch the magnet thesis directly. But it is one of the seven medium and heavy rare earths under the April 2025 licensing regime, the carve-out this page has repeatedly noted was never suspended by Announcement No. 70, and it is now demonstrably being used as country-specific leverage. It lands on the venue where four of the six tier-one names trade, and on the chip-equipment makers that led Tokyo down the same morning.
Ambarella’s mystery third bidder has acquired a name, and still not a source. Ambarella rose 4.40% to $66.00 on 2.38 times average volume, the heaviest session in the name this month. Reports dated 21 October named Nvidia as the third party evaluating the company, said to be working with Goldman Sachs, joining the 31 July FT report of NXP talks and the 9 October Qualcomm speculation. The sourcing is the same as it was on 20 October when this page last recorded it: no named outlet, no filing, no company confirmation, and one of the two pieces carrying it reported the session at about 7% intraday while the stock settled at 4.40%. The sector did part of the work regardless, with the semiconductor index up 1.97% and Nvidia up 2.54% the same day. Priced on the tape rather than the rumour, Ambarella is the only tier-one name whose drawdown narrowed this run, and at 31.7% below its high it is now the tightest on the page.
Two base rolls, pulling the two tiers in opposite directions, and one of them undoes last run’s headline exactly. The Western one-year base rolled from 20 onto 21 October 2025, which takes 20 October 2025, the session the market traded MOFCOM’s extraterritorial controls, back out of the twelve-month window. Last run this page recorded that same session rolling in and costing four minerals cells four to six points with no price behind them; rolling out, it hands back more than it took. Perpetua −31.4% to −21.0%, of which 10.3 of the 10.4 points is the base; Lynas −42.0% to −32.9%, 9.0 of base; USA Rare Earth −68.2% to −61.7%, 5.9 of base; MP −52.9% to −46.7%, 5.7 of base. Not one of those four moved as much as half a per cent on the day. Tokyo’s base rolled 21 onto 22 October 2025 and ran the other way, but there it was swamped: of the 16.5 points Harmonic Drive’s 1y cell lost, 15.8 is the session and 0.7 the base. The four-week bases also split by venue for the first time in three runs, Tokyo pinned on 18 September by the 21–23 September holiday block while the Western names rolled to 21 September; that is what takes MP’s 1m cell from −17.4% to −21.0% on a session it rose. Not one of the eleven 52-week highs rolled, so the whole drawdown column this run is price.
The correlation window has not moved, so nothing in section 04 is new, and that is the right answer. ISO week 43 does not close until Sunday 25 October, so the sample holds at 104 aligned weekly closes and 103 returns from week 2024-43 to week 2026-42, a fourth consecutive run on the same window. Both bucketing methods were rebuilt from two years of daily closes and agree to four decimal places on all eleven pairs and on the headline: basket-to-REMX 0.3885, comfortably inside the 0.30–0.50 band, so the relationship is not news this run even though the tape was. The four-week comparison did move, and it narrowed again: REMX is 24.1% lower against the basket 16.7% lower, a gap of 7.4 points from 8.3. Unlike last run it is not clean, because the Western bases rolled and Tokyo’s did not: on a common 18 September base the gap would read 6.5, so price narrowed it 1.8 points and the calendar widened it 0.9. A 2.65% Tokyo rout narrowing the gap to a rare-earth ETF is a reminder of what the 0.39 actually means: the basket fell hard on a Japanese AI de-rating that had nothing to do with magnets at all.
21 Oct 2026. Nidec published the redacted Report, on the day it said it would, and the number in it is ¥27.6bn. “Nidec Discloses the Investigation Report Issued by the Executive Responsibility Investigation Committee”, dated 20 October in Kyoto, says the redaction is complete and attaches the Report. The trigger this page has pointed at for ten weeks has now fired in full: receipt on 16 October, the Committee’s conclusions and the nine names the same day, and the Report itself on 20 October. Two things in the release are new and neither was in the 16 October items. The appropriation of surplus was excluded from the Committee’s investigation, because Nidec had not filed amended securities reports for past fiscal years and the distributable amount that would have been examined is therefore uncertain; the company says it will examine legal action on that question separately, once the restatements are done. And only a Japanese digest is disclosed, with the English version promised and not yet published — which is why the English-language wire route this page normally reads carries nothing, and the figures reach English readers through Jiji, Bloomberg and Nikkei reading the Japanese text. Those three agree: damage of about ¥27.6bn, made up of third-party investigation costs, audit costs and the Tokyo Stock Exchange penalty; founder Shigenobu Nagamori and former president Hiroshi Kobe found in breach of duty in most of the matters examined; and the apportionment among the nine left to a court, because the Report says contribution is too hard to split. Nidec itself is still only “deliberating legal actions such as lodging a claim for damages”, so the board resolution an aggregator reported on 16 October — which this page’s audit marked wrong at the time — still has not happened. Seven days remain to the 28 October anniversary.
The tape still will not trade it, and it has not traded it on any session since the Committee reported. Nidec closed ¥1,762, up 0.23%, four yen, on 8,954,900 shares against a twenty-session average of 17,266,920, or 0.52 times normal. It traded as low as ¥1,748, the lowest print on this page’s record and the lowest intraday since 18 April 2016, and the ¥1,758 it closed at the session before stands as the lowest close since 12 February 2016. A company report putting a ¥27.6bn number on its founder’s liability moved the close two tenths of a per cent on half the usual turnover. The reading this page has made throughout holds: the bid going away rather than a seller arriving. Nothing is dated 21 October on Nidec’s feed.
Tokyo rebounded and the component tier led it, with one exception. The Nikkei closed 64,621.24, up 545.82 points or 0.85%, confirmed against the 21 October daily bar, regularMarketPrice stamped 15:45 JST and Nikkei Inc.’s own index page. Harmonic Drive +2.47% to ¥5,390 was the largest rise on the page in Tokyo, on 1.32 times average volume, a second consecutive session above its own average. TDK +2.42% to ¥2,902.5 on 0.74 times recovered most of the previous session’s fall and leaves it the only name anywhere on this page with a positive four-week cell, at +2.0%. Nabtesco +1.49% to ¥4,357 on 0.80 times took back two fifths of the largest fall on the page the day before, and Aichi Steel +0.42% to ¥2,879 on 0.75 times was the smallest move in Tokyo. The exception is Yaskawa, −1.71% to ¥4,204 on 1.06 times average volume — the only tier-one name in Tokyo to fall, its lowest close since 31 March 2026, and the only Tokyo line here to trade above its own average while doing it. No company on this page filed anything dated 21 October.
Tuesday in New York was a breadth session, and the minerals tier finally rose into one. The Dow rose 1.02% to 52,495.88 and the Russell 2000 1.65%, while the S&P 500 added 0.28% to 7,712.57, the Nasdaq Composite fell 0.02% and the PHLX Semiconductor index 0.29%: small caps and the average stock, not the megacaps. The tier went with the breadth rather than the index. Lynas +3.35% to $8.03 on 1.14 times average volume was the largest rise anywhere on the page; Compass Diversified +3.08% to $10.36 on 0.78 times was the second, the mirror of a session before that put 2.15 times the volume behind four tenths of a point; Schaeffler +2.08% to €5.89 was the best session in tier one and pulls its drawdown, still the widest there, in to 50.8%. REMX +1.31% to $54.06 on 0.89 times average volume is its best session since 21 September and the first up-day of any size in this account. The two largest producers in it barely moved, MP +0.26% to $39.05 on 0.68 times and USA Rare Earth −0.30% to $10.06 on 0.70 times, extending its lowest close since 9 July 2025 by three cents; Perpetua −0.27% to $18.72 and Novonix −2.31% to $2.11, a fresh record closing low on exactly its average turnover, went the other way. So the fund rose more than four of the five producers it holds and less than one, which is the ninth session in this account where it and its largest holdings disagreed and the first of them in the fund’s favour. Nor did the tape hand it the move: gold fell 0.54% and the gold miners 0.43%. Nothing on the page moved four per cent either way.
One lookback roll did more to this page’s columns than any price did, and it is the same calendar day pulling two columns in opposite directions. The Western one-year base rolled from 17 onto 20 October 2025, the session the market traded MOFCOM’s extraterritorial controls and the Western producers re-rated on scarcity. Four minerals 1y cells fall four to six points on it with no price behind them: Lynas −36.4% to −42.0% on a base rolled $12.21 to $13.84, a 13% one-day step, on a session the ADR rose 3.35%; USA Rare Earth −63.6% to −68.2% on $27.73 to $31.59; Perpetua −25.9% to −31.4% on $25.32 to $27.27; MP −51.8% to −52.9% on $80.79 to $82.90. The same day leaving the other window did the opposite to MP’s drawdown: its 52-week high rolled $80.89 to $76.80, taking the cell from 51.9% to 49.2% and under fifty per cent for the first time in this account, on a quarter-point session. Held at today’s price on the old high the cell would read 51.7%, so of the 2.7 points, price put in 0.2 and the base roll 2.5. Novonix’s high rolled $21.50 to $20.70, which this time nearly cancelled against a 2.31% fall, leaving 89.8%. Tokyo’s own base moved one session, 20 onto 21 October 2025, and it was small: Harmonic Drive’s 1y adds a point to +65.6% on a base rolled ¥3,200 to ¥3,255, and Yaskawa’s improves to −3.4% on a session it fell. Read the bases before reading any of those nine cells.
The four-week bases did not move at all, and the correlation window has not either, so one figure on this page moved for a clean reason. All seven four-week bases sit on 18 September for a second consecutive run, the first time this series has had two, so the tier gap narrowing from 8.6 points to 8.3 is price with no calendar in it: REMX’s four-week cell went −22.7% to −21.7% and the basket’s −14.2% to −13.4%, both tiers up and the minerals side up more. ISO week 43 does not close until Sunday 25 October, so the correlation sample holds at 104 aligned weekly closes and 103 returns from week 2024-43 to week 2026-42, where the 17 October run set it. Rebuilt from two years of daily closes and cross-checked against weekly bars bucketed by exchange-local ISO week, both methods again give basket-to-REMX at 0.3885, so the headline 0.39 stands and every correlation cell on the page is unchanged from the corrected figures the last run published. The next window advance is the week-43 sample on the 26 October run.
A wire put Tokyo’s session half a point wrong, and the audit now carries that error class three times. A report dated 21 October headlines the Nikkei “up 0.97%” and its index table reads 64,696.00, up 620.58 points. The index closed 64,621.24, up 545.82, or +0.85%. The two figures are internally consistent with each other and with the correct previous close, so this is not a transcription slip but a different print — one that sits inside the day’s range of 64,558.60 to 65,404.63 and is not the 15:30 auction. It joins the “Nikkei −2.64%” row from 19 August and the “down 1.16%” row from yesterday. Check the index against the index.
Three smaller items, none of them a session. The Schaeffler daily bar went null again, this time for 20 October, and this one is worse than its predecessors: the volume field is null as well as the close, so the twenty-session comparison had to be made against the quote-level volume. Two routes agree on €5.89 and the continuous five-minute tick would have given €5.94, a +2.95% session rather than the +2.08% one that happened, so the rule earned its keep in the opposite direction from last week. That is the thirtieth hole; 16 October has backfilled at exactly the €5.83 published here. On the policy clock, nothing has changed and the date is closer: MOFCOM Announcement No. 70 still expires 10 November 2026, 20 days away, with no Chinese instrument extending it and the separate gallium, germanium and antimony suspension running to 27 November. And three open items stay open: USA Rare Earth has still not scheduled third-quarter results with the 600 mtpa Stillwater run-rate and first magnet invoice both guided to this quarter, Lynas’s September-quarter report is still due late October, and Perpetua’s $2.9bn EXIM facility is still board-approved rather than closed.
20 Oct 2026. The only rare-earth expiry written into Chinese law is still 10 November, and this page’s own audit said otherwise. MOFCOM Announcement No. 70 of 2025, dated 7 November 2025, suspended Announcements 55 to 58, 61 and 62 until 10 November 2026. Two statements point past that date and neither is a Chinese instrument: Treasury Secretary Bessent said on 23 September that the Busan Agreement would run to 10 January 2027, and a ministry official said the same on 28 September. Trade-policy trackers reading MOFCOM’s own 2026 announcement list, one on 1 October and one on 16 October, found it running to Announcement No. 48 with nothing extending, replacing or amending No. 70. So the cliff edge this page recorded as removed is removed in a US characterisation of the truce and not in the published text, and the date in the text is 21 days away. The two carve-outs are unchanged and both are on the page already: the April 2025 licensing regime covering seven medium and heavy rare earths was never suspended, and the separate suspension covering gallium, germanium, antimony and superhard materials runs to 27 November 2026. The claim audit now carries this against this page’s own “Wrong” verdict on the 6 October piece, which was too strong: the headline date it counted down to is the only one Beijing has published.
Nidec said it would disclose the redacted Report on 20 October “or so”, and as of this run it has not. Its IR feed carries four October items: two dated 16 October, the receipt and the Committee’s conclusions, plus one on 2 October and one on 1 October. Nothing is dated 19 or 20 October. Tokyo has settled and the feed can still post after hours, so this is a status rather than a miss, and the test is whether it appears before the next run. The tape stayed where it has been: Nidec closed ¥1,758, down 1.73%, a third consecutive new low beneath the ¥1,789 that was itself the lowest close since 11 July 2016, on 7,322,600 shares against a twenty-session average of 17,046,995, or 0.43 times normal. Three sessions of fresh lows on 0.72, 0.32 and 0.43 times average turnover is the reading this page has made throughout: the bid going away rather than a seller arriving. Eight days remain to the 28 October anniversary.
Tokyo fell across the board and the published percentage is wrong in the wires again. The Nikkei closed 64,075.42, down 841.28 points or 1.30%, confirmed against both the daily bar and regularMarketPrice stamped 15:45 JST. One wire put the session at −1.16% and an historical table at 64,224.00; neither is the close, and the audit carries the first as the same error class as the “Nikkei −2.64%” row. Decliners beat advancers 2,821 to 710. Underneath, the gearbox layer took the worst of it: Nabtesco −3.70% to ¥4,293, the largest fall anywhere on this page today and its lowest close since 29 July 2026, on 0.78 times average volume; TDK −2.71% to ¥2,834, its lowest since 17 September, on 0.68 times; Aichi Steel −2.35% to ¥2,867 on 0.74 times; Yaskawa −2.20% to ¥4,277 on 0.84 times. Harmonic Drive was the exception at −0.75% to ¥5,260, and the only Tokyo line here to trade above its own average, at 1.24 times. Nothing on the page moved four per cent, and no company on it filed anything dated 20 October.
Monday in New York went the other way, and the minerals tier fell into a rising market. The S&P 500 rose 0.41% to 7,690.73, the Nasdaq Composite 1.11% and the PHLX Semiconductor index 1.61%, recovering part of Friday’s chip-led fall. The tier sold anyway: USA Rare Earth −4.27% to $10.09, the largest Western fall on the page and its lowest close since 9 July 2025, on 0.74 times average volume; MP −3.04% to $38.95, its lowest close since 29 July 2026, on 0.72 times; Perpetua −2.29% to $18.77 on 1.28 times; and REMX −1.60% to $53.36, its lowest close since 7 August 2025, on 0.89 times. That is the fund falling less than the two producers inside it for the eighth time in this account, and the explanation available on earlier sessions is not available here: the equity market was up, not down. Nor is it a metals session. Gold fell 0.61% and the gold miners 1.41%, so Perpetua at roughly 1.6 times the miners is the documented amplification of a metals move rather than an antimony signal, which is the check that card exists to make. Two names rose: Lynas +1.83% to $7.77 on 1.29 times average, its best session since 15 October, and Novonix +0.47% to $2.16 on 0.59 times, off the record low it set on Friday. Compass Diversified +0.40% to $10.05 came on 2.15 times average volume, the heaviest relative turnover on the page, with nothing on its feed dated to it.
Ambarella rose 1.72% to $62.76 on 1.76 times average volume, and its own sector did most of that. A third reported bidder entered the sales process last week, after the 31 July Financial Times report of NXP talks and the 9 October Qualcomm speculation. The October items trace to unverified “uncooked” market alerts repeated by aggregators, and neither Ambarella nor any named party has filed or said anything, so they are reported here as circulating talk and not as a process. The session itself does not need them: the stock rose 1.72% on a day its semiconductor index rose 1.61%, which is sector beta and not a deal premium. This card’s standing read was that Ambarella’s falls are sector beta with the robotics line doing none of the work; the same test applied to a rise gives the same answer.
The one-year lookbacks rolled, and they did more damage to this page’s columns than any price did. Tokyo’s 1y base moved from 17 October 2025 to 20 October 2025 and the Western names’ from 16 to 17 October 2025, off a weekend into a higher session. Harmonic Drive’s 1y cell falls from +75.2% to +64.4% on a base going ¥3,025 to ¥3,200, with the price down eight tenths of a per cent: ten of the eleven points are the calendar. Yaskawa’s turns negative, +4.8% to −4.4%, on a base going ¥4,173 to ¥4,474. TDK’s goes +29.2% to +19.1% on ¥2,255.5 to ¥2,379, and Aichi Steel’s YTD flips negative, +0.5% to −1.9%, on a 2.35% session. Two 52-week highs rolled as well, and both flattered a drawdown cell: MP’s from $83.99 to $80.89, which improved its drawdown from 52.2% to 51.9% on a session the stock fell three per cent, and Novonix’s to $21.50, which took its cell from 91.0% to 90.0% on a price that rose half a per cent. None of those four cells is a judgement about the business; read the base before reading the number.
All three venues print a new close and the four-week bases have converged again, this time on a weekday. Tokyo reads 20 October, Xetra and the US 19 October, so the structural one-session gap is back to its normal shape rather than closed. All seven four-week bases nevertheless sit on a single date, 18 September, because the Western lookback rolled onto the same Friday Tokyo’s already used: the second single-base run in this series and the first that is not a weekend refresh. The tier gap widened from 7.7 to 8.6 points, REMX −22.7% against the basket −14.2%, and for once the decomposition cuts against the headline. Held at today’s prices on the old 16 September Western base the gap would be 9.6 points, so the base roll took 1.0 point out and price put 1.9 points in. The widening is price, and the calendar worked against it, which is the opposite of the last three runs.
The correlation window has not moved, but three published cells were slightly off and are corrected. ISO week 43 is still in progress and does not close until Sunday 25 October, so the sample holds at 104 aligned weekly closes and 103 returns from week 2024-43 to week 2026-42, exactly where the 17 October run set it. Rebuilt from two years of daily closes and cross-checked against weekly bars bucketed by exchange-local ISO week, both methods give basket-to-REMX at 0.3885, so the headline 0.39 stands and the relationship has not changed. Three cells recompute one point higher than the figures carried since 17 October: Lynas to REMX 0.6471, published 0.64; Nidec to the basket 0.6229, published 0.61; Harmonic Drive to the basket 0.8152, published 0.81. On an unchanged window that cannot be the market moving, so it is a carried rounding error rather than news, and the bar charts and the Lynas table cell are corrected to 0.65, 0.62 and 0.82. The audit carries it as this page’s own. The next window advance is the week-43 sample on the 26 October run.
Four smaller items, none of them a session. Schaeffler’s 13 August 2026 release, which this page had not carried, says it has completed validation testing of formed strain wave gearboxes for humanoid robots and plans mass manufacture from 2027, starting in Germany, claiming more than 25% lower manufacturing cost and more than 75% lower material consumption. Those are company claims and not verified here, but the product is Harmonic Drive’s own, so the parts table now names Schaeffler on the strain-wave row with its 2027 date attached. On the magnet input, the China Rare Earth Industry Association posted its price index at 263.6 on 9 October 2026, against 267.0 on 4 August, with NdPr oxide at ¥729.8–749.8/kg, which brackets Benchmark’s 25 September assessment of RMB729,500/t rather than contradicting it. USA Rare Earth has not scheduled third-quarter results, and both the 600 mtpa Stillwater run-rate and first magnet sales are guided to this quarter, which has ten weeks left. Lynas’s September-quarter report is due late October, and Perpetua’s $2.9bn EXIM facility is still board-approved rather than closed, with the company guiding documentation and close to the fourth quarter.
19 Oct 2026. The Committee’s conclusions are on the record with nine people named on them, and Tokyo traded a third of its normal volume. Nidec’s second release of 16 October, “The Investigation Report Issued by the Executive Responsibility Investigation Committee”, carries the Committee’s conclusion that the following “are believed to be liable in damages”: Shigenobu Nagamori, Hiroshi Kobe, Akira Sato, Hiroyuki Yoshimoto, Jun Seki and Mitsuya Kishida under Article 423, Paragraph 1 of the Companies Act of Japan; Kazuya Murakami and Hiroyuki Ochiai under the same article; and a former vice president, identified only as a former CFO of the company from May 2022 to March 2026, under either Article 644 and Article 415, Paragraph 1 of the Civil Code or Nidec’s own Vice President Regulations. That is nine individuals, eight named and one not, with the caveat that “the extent of damage should be determined based on the degree of their contribution”. The Report itself is still being redacted: Nidec “plans to disclose the Report on Tuesday, October 20, 2026 or so”. Nine days remain to the 28 October anniversary.
This page said on 17 October that nothing had posted, and a release dated the day before said otherwise. The 17 October run recorded that “neither the redacted Report nor the promised actions has posted”. The release above is dated 16 October and sits on the same disclosure feed, one item above the receipt this page did carry. It is not the redacted Report, which is still pending, but it is the Committee’s conclusion and the names, which is the substance the run said was absent. The miss is the same mechanism the audit recorded against the Oasis letter and the Schaeffler order book: a dated company item on the public record, not found because the feed was read for one item rather than listed. The claim audit carries it as this page’s own.
The tape did not trade it. Nidec closed ¥1,789, down 0.61%, its lowest close since 11 July 2016, on 5,439,500 shares against a twenty-session average of 16,858,410, which is 0.32 times normal and the thinnest session of this entire episode. It traded as low as ¥1,749, which is a new 52-week low and the lowest print on this page’s record, and as high as ¥1,803. A board committee naming the founder and eight others as liable in damages moved the close six tenths of a per cent on a third of normal turnover. On 15 October this page noted that a ten-year low on below-average volume is the bid going away rather than a new seller arriving; at 0.32 times, neither side turned up.
Two figures circulating on this session are wrong, and both are now in the audit. An aggregator has Nidec’s board “resolving on October 16 to pursue litigation against nine former executives”. Nidec’s own release says it “will consider necessary legal actions such as filing a claim for damages”, which is deliberation and not a resolution, and the nine are not all former executives: Murakami and Ochiai are carried without that qualifier and one is a former vice president rather than a director. Separately a wire has the stock “down 2.8% to 1,760.0 yen” on Monday. The close was ¥1,789, down 0.61%; ¥1,760 is an intraday print from a session whose low was ¥1,749, and −2.8% from Friday’s ¥1,800 is ¥1,749.6, which is the low rather than the close. This is the same error class as the “Nikkei −2.64%” figure the audit already carries.
04 The REMX Question (Rare Earth ETF)
Every humanoid needs roughly 1.3 kg of NdPr magnet material, and China makes about 90% of the world's rare-earth magnets. That is a real physical dependency. It does not follow that a rare-earth ETF is a way to own robotics — so I tested it rather than assuming it.
Correlation with REMX
Weekly return correlation, 103 weekly returns from 104 aligned weekly closes to 16 Oct 2026 (ISO week 42, the last complete week: week 43 opened with the 19 October Tokyo session and does not close until Sunday 25 October, so the in-progress week is dropped and the window is unchanged for a fourth run; three cells carried on it were corrected on 20 October, see the method note), re-bucketed from daily closes into true ISO calendar weeks so Tokyo, Xetra and New York are matched to the same week — not Yahoo's native weekly bars, which anchor Tokyo/Xetra weeks about one week off from New York's.
Robotics component names Rare earth & critical minerals names
The two groups trade differently, but less starkly than this page showed before the 18 Aug alignment fix. Lynas, MP Materials and Perpetua move with REMX at 0.65, 0.56 and 0.53. The robotics names sit between 0.18 and 0.46 — every one of them, Nidec included, is a genuine positive correlation once Tokyo and Xetra weeks are properly aligned to REMX's calendar. A weekly-rebalanced equal-weighted robotics basket correlates to REMX at 0.39, not the 0.19 this page carried before the fix — and the basket definition is worth stating, because it moves the figure: held without rebalancing the same six names give 0.40, which is the construction this page published through 16 September without saying which one it meant. Both are defensible; only one was disclosed. See the method note and the claim audit.
Meanwhile those same robotics names correlate to the basket at 0.55–0.82. They still form a tighter bloc with each other than with REMX, which is a real but partial exposure, not a stand-in. The four-week columns still disagree, but by much less than they did: from 18.6 points to 14.9, then 12.6 and 10.9, back out to 13.0 and 13.2, sharply in to 8.8 and 8.4, out again to 9.8, in to 7.7, out to 8.6, in to 8.3, and now in again to 7.4 points. REMX is 24.1% lower on the four-week lookback the tables use, against the basket 16.7% lower, the mean of the six tier-one names’ own 1m cells, which is the construction, stated, because this page once published a basket four-week figure of 5.2% that cannot be rebuilt from any of them. And this run the calendar is back in it, on one side only. The four-week bases have split by venue: Tokyo’s stay pinned on 18 September for a third consecutive run, because the exchange was shut on 21, 22 and 23 September, while the Western names’ rolled to 21 September. REMX’s cell went from −21.7% to −24.1% on a session it fell 1.28%, and the basket’s from −13.4% to −16.7% on the 2.65% Tokyo rout, for a 7.4-point gap of which 1.8 points of narrowing is price and 0.9 of widening is the base roll. The previous run was the clean one: both tiers up, the minerals side up more, and no base anywhere to argue with. REMX’s cell cannot have moved: it has no new close, and sits unchanged at −19.8%. The basket’s went from −10.0% to −12.1%, and those 2.0 points split 0.3 to price and 1.7 to the calendar: Tokyo’s base rolling off 16 September lifted Nidec’s own four-week base from ¥2,646 to ¥2,856 and its 1m cell from −32.0% to −37.4%, which alone is most of the move in a six-name mean. So five sixths of the narrowest gap on record is a date change, and one run after this section finally got to describe a clean one. Read the bases before reading convergence. The decomposition that follows belongs to the earlier run in which the gap went the other way, and is kept as the record of it. Both sides fell and one fell much harder: REMX’s cell went from −13.6% to −17.5%, 3.9 points worse on its 4.58% session, while the basket’s went from −2.7% to −4.6%, 1.9 points worse. The difference between those two, 2.1 points, is the whole of the widening, and all of it is price. That matters because of what this section has had to say three runs in a row: the narrowing from 18.6 points to 10.9 was mostly the REMX base rolling off a crash session rather than either tier moving toward the other, and the one run where a real price did the work, Ambarella’s seven-point rumour session, pushed the gap the wrong way. Now a run with no base movement at all has pushed it wider again, and the mechanism is the obvious one: a minerals tier having its fourth day of rare-earth-specific selling, against a robotics tier that fell on a Tokyo de-rating and partly offset it with Yaskawa and Aichi Steel. Three runs of base-driven convergence have been undone by one session of genuine divergence. Read the bases before reading convergence, and read the prices before reading this. What survives is the point this section keeps having to make. The two tiers arrive at four-week numbers by different routes: REMX through rare-earth-specific selling that ran two sessions on a tape where gold was bid and then stopped, and the robotics names through a Tokyo de-rating of the gearbox layer, a partial reversal of it, and a US semiconductor name moving on an unverified bid report. A correlation of 0.39 permits exactly this. Equal four-week returns are not shared exposure, and unequal ones are not divergence; that is the error this whole section exists to refuse. And this is the run the window finally moved, on the date this note named in advance. Week 41 closed on Sunday 11 October, so the sample now ends at the week ended 9 October and the start rolls forward with it. The prediction was that one of the busiest weeks this page has recorded would land on the correlation as very little, and it did: basket-to-REMX goes from 0.3681 to 0.3687, a move of seven ten-thousandths. The four named check figures all came in as described, which is worth recording because they were published before the week entered the sample: Harmonic Drive entered at ¥6,220 against ¥6,570, a 5.33% weekly fall, TDK at ¥3,268 against ¥3,198, a 2.19% rise, Yaskawa 0.48% lower, and Ambarella at $67.43 against $68.65, 1.78% lower on the week despite Friday’s seven-point session.
But the reason the figure barely moved is not the one the word “cancels” implied, and the distinction is the whole lesson of this run. Week 41 did not cancel inside the basket. It was a broadly down week in both tiers: the basket fell 3.27% and REMX 3.54%, with five of the six tier-one names lower (Schaeffler 5.62%, Harmonic Drive 5.33%, Nidec 5.05%, Ambarella 1.78%, Nabtesco 1.36%, Yaskawa 0.48%) and four of the five minerals names lower (USA Rare Earth 7.58%, Novonix 5.02%, Lynas 2.78%, MP 1.77%, against Perpetua up 1.28%). A week in which both sides fall by almost exactly the same amount is not a quiet observation, it is an observation sitting almost on top of the existing relationship, so it confirms the correlation rather than changing it. That is a different fact from the one this note predicted, and the predicted number was right for the wrong reason. Had the two tiers fallen by materially different amounts the same busy week would have moved the figure hard in one direction or the other.
The one name the new week did move is Schaeffler, and it moved on both axes. Its correlation to the basket rises from 0.5997 to 0.6070 and to REMX from 0.4564 to 0.4645, the largest single change on the page when the window last advanced, and the only one big enough to round differently, from 0.60 to 0.61. It earned that by being the worst line in the basket in the week the basket fell, which is what a high-beta member does to its own correlation. Three other cells round differently on much smaller moves: Nabtesco to REMX from 0.22 to 0.21, Yaskawa to the basket from 0.82 to 0.81, and MP to the basket from 0.19 to 0.20; in the magnet tier TDK goes from 0.52 to 0.51 and Aichi Steel from 0.17 to 0.18. Everything else is unchanged at two decimal places: Lynas 0.6333, MP 0.5524, Perpetua 0.5248, USA Rare Earth 0.3678, Novonix 0.3603. The robotics names now run 0.18 to 0.46 against REMX, and 0.55 to 0.82 against the basket. On the 16 October window five cells round differently again, all by a single point and none of them changing a reading: Lynas to 0.64, MP to 0.56, Perpetua to 0.53, Novonix to 0.37 and USA Rare Earth to 0.38, with Nabtesco back out to 0.24 and Nidec to 0.35 against REMX. No conclusion in this section moves, which is the honest outcome of advancing a 103-week window by one week, even a week with this much in it. Separately, the four-week columns did move this run, and for once without any help from the calendar: the gap widened from 10.9 to 13.0 points on price alone, with all seven bases still sitting on 11 September. See the decomposition above.
Part of the reason is what REMX actually holds. Roughly 24% of the fund is in rare-earth companies and about 38% in lithium names — more lithium by weight than rare earths, as of April 2026. Top holdings run to MP Materials, Lithium Americas, Lynas, China Northern Rare Earth and Albemarle.
So a bet on REMX is substantially a bet on the lithium cycle and on Chinese miners, with roughly a third of the fund in Chinese companies. It is weighted to miners rather than to the separators and magnet makers a robot actually consumes.
What REMX is really tracking. Its drivers are Chinese export policy, MIIT production quotas and the NdPr price — not humanoid order books. Chinese Pr–Nd oxide was assessed at $97.40/kg on 4 August 2026, down about 12% from July and still below the $110/kg floor set under the US DoD–MP price-protection agreement; China's rare-earth price index closed the same day at 267.0, well below the early-2026 peak near 310. What has moved the ETF since is policy, not price: an upsized US government commitment behind USA Rare Earth's Brazilian offtake, and Beijing's own 2026–2030 materials strategy set out at a State Council briefing on 26 August. What has moved it over the past fortnight is neither, and the four-session run that closed September has now ended. REMX fell 5.1% to $72.42 on 10 September, its worst session since 23 June, inside a broad metals liquidation: US producer prices a tenth above forecast, a light Treasury buyback, the ten-year at 4.94%, silver −5.3% and the gold miners −3.5%. Then on 11 September the macro reversed — the S&P rose 0.86%, gold 0.6%, silver 1.1%, the gold miners 1.1% — and REMX fell anyway, 2.6% to $70.56. That second session is the tell. A liquidation accounts for the first day; only something specific to rare earths accounts for the second. Three smaller falls followed, 2.3% to $68.93 on 14 September alongside the whole precious complex, 0.9% on 15 September and 1.0% to $67.61 on 16 September — the lowest close since 31 July, and the sixth down session running. Then the run ended, and policy is what ended it: +2.2% to $69.10 on 17 September, flat at $69.06 on 18 September, and +1.8% to $70.27 on 21 September, when Trump's 18 September security agreement with Denmark and Greenland reopened the trade in anything holding an Arctic deposit. Greenland-linked micro-caps moved in multiples that session on press accounts not verified here, while the ETF moved 1.8% — roughly the difference between owning the story and owning a diversified basket. A third rise followed, +1.0% to $70.97 on 22 September, and it was the weakest of the three: the whole minerals tier was bid that session but none of it led, and the day's largest mover on this page was a gold name moving with gold miners rather than with rare earths. Then the run ended on 23 September, −3.1% to $68.76 — and the size of it is the point. Hawkish Fed remarks put the dollar index back above 100 and liquidated the whole metals complex that session, taking the gold miners down 4.4% and silver 4.2%; REMX fell less than either. The rare-earth-specific headline of the day — Bessent extending the Busan Agreement from 10 November to 10 January, removing a cliff edge that had been part of the scarcity premium — shows up not in the ETF but in the single names that ran past it, USA Rare Earth −6.8% and MP −4.2%. Then 24 September extended it without the same excuse: −2.7% to $66.93, its worst session since 10 September and its lowest close since 31 July, on a day the dollar index merely held above 101 rather than surging and the gold miners fell 1.3% — a third of REMX's move. So this one is not a dollar session and not a precious-metals session; it is rare-earth selling on its own account, and the two largest names in the fund did not lead it, MP closing higher. Then 25 September took it a third session lower and barely at all, −0.4% to $66.66, extending the lowest close since 31 July on no rare-earth news of any kind — this page called it the fourth at the time, and the claim audit carries the correction. The fourth came on 28 September and came hard, −3.3% to $64.47, its lowest close since 6 November 2025, on a day whose cause was an oil price and a rates repricing rather than anything upstream of a magnet. Then the run ended on 29 September, +0.5% to $64.79, and the way it ended is the more interesting half: every US rare-earth and minerals name on this page fell that session, MP 2.2%, USA Rare Earth 2.4% and Lynas 1.0%, while the fund holding them rose. A diversified basket can go up on a day its rare-earth constituents go down, which is the whole of this section in one session. Then 30 September ran the same experiment backwards and got the mirror answer: REMX fell 0.8% to $64.26, extending the lowest close since 6 November 2025, on a session in which MP rose 4.1% and Lynas 1.3%. Two consecutive sessions in which the fund and its largest holdings moved in opposite directions, once each way, is a better demonstration of a 0.37 correlation than any single day could be. Two sessions since have added almost nothing either way: −2.0% to $62.95 on 1 October, its lowest close since 23 September 2025, on the session Lynas fell nine per cent on its own announcement, +0.4% to $63.21 on 2 October, a fund barely moving while two of its holdings rose two and a half per cent and a third fell one and a half, +0.5% to $63.54 on 5 October, a third consecutive session in which the fund and its largest holdings disagreed about direction — MP, Lynas and Novonix all fell while the fund that holds them rose — and +0.4% to $63.79 on 6 October, which makes the same point by magnitude rather than by direction: MP rose 3.6% and Lynas 2.3%, and the fund holding both of them added four tenths of a point. Then 7 October ended the disagreement and ended it downwards: −3.0% to $61.88, the lowest close since 23 September 2025, and this time every minerals name on this page went with it — MP −4.6%, USA Rare Earth −3.7%, Perpetua −3.1%, Novonix −2.5%, Lynas −1.1%. After three sessions in which the fund and its largest holdings disagreed about direction, one in which they all agreed is the less interesting outcome statistically and the more informative one here, because of what was not moving: the S&P 500 fell 0.2% that session and the gold trust 1.7%. A fund that falls thirteen times its equity index on a day the broad market is flat is not being liquidated along with everything else; it is being sold on its own account. No rare-earth disclosure of any kind is dated to it, and no company on this page announced anything. That is the fourth distinct cause this section has had to record in a month, after a metals liquidation, a security treaty and a dollar move. Then 8 October ran it a second day, −2.1% to $60.58, the lowest close since 12 September 2025, and this session closes off the one explanation that was still available. On 10 and 23 September this page could attribute rare-earth selling to a liquidation of the whole metals complex. Not here: gold rose 0.73% and the gold miners 1.47% while REMX fell, with the S&P 500 down 0.47%. A fund that falls two per cent on a day the precious complex is bid and the broad market is flat is not being liquidated alongside anything; it is being sold on its own account for a second consecutive session. The single names went with it, USA Rare Earth −4.5% to its lowest close since 31 December 2025, Lynas −2.9% to its lowest since 2 January 2026 and Novonix −2.8% to a new two-year low, while MP sat almost still at −0.4%. Two causes are circulating for it and neither is dated to the session: August customs data showing 512 tonnes of Chinese magnets shipped to the United States, released 20–21 September, and the Xi–Trump state visit of 23–25 September. Both are standing conditions of this tier rather than events in it, and the next scheduled meeting is APEC on 18–19 November. Then 9 October ended the run, and the way it ended is the useful part: +0.6% to $60.97, a rise into a session in which the whole precious complex was bid, gold +1.57% and the gold miners +2.95%, with the S&P 500 up 0.59%. A fund that adds six tenths of a point on a day the gold miners add three is not leading a recovery in rare earths; it is declining to fall while everything adjacent rallies. Its own constituents split underneath it, Lynas +1.5%, Novonix +3.2% and Perpetua +2.6% against USA Rare Earth −0.6%, which extended its lowest close since 31 December 2025 on the day its tier recovered, and MP exactly unchanged at $46.14. Then 12 October resumed it and took the fund lower than any session in this whole account: −4.6% to $58.18 on 1.18m shares against 571,200, roughly double the volume, closing within a quarter of its low, and its lowest close since 10 September 2025. This time the constituents did go with it, and hard: MP −6.0%, USA Rare Earth −5.9%, Lynas −3.1%, Novonix −3.1%, Perpetua −1.9%. And the adjacent complex again refuses to explain it: gold fell 0.63% and the gold miners 0.84%, a fifth and a seventh of the fund’s move, so this is neither a dollar session nor a precious-metals liquidation. It is a fourth distinct day of rare-earth selling on its own account, with no company disclosure and no policy item dated to it. Two sessions since have taken it lower still, −2.91% to $55.78 on 14 October, its lowest close since 8 August 2025 and again a smaller fall than every producer it is meant to proxy. Then 15 October made the same point with the sign reversed: the fund closed −0.07% at $55.74, effectively unchanged, on 2.75 times average volume and after an intraday low of $53.75, while MP rose 4.16% and USA Rare Earth 7.46%. Then 16 October resumed the fall and took it through the August 2025 floor: −2.71% to $54.23 on 1.34m shares against a twenty-session average of 753,045, which is 1.78 times average, and the lowest close since 7 August 2025. This time it did not fall less than its constituents, it fell more than most of them — MP −3.46%, USA Rare Earth −4.96%, Novonix −1.38% and Lynas −0.52%, with Perpetua up 3.75% — which after seven runs of the opposite complaint is worth recording as the exception rather than a change of character. Then 19 October put the character straight back and removed the last macro excuse available to it: −1.60% to $53.36 on 702,000 shares, 0.89 times a twenty-session average of 788,965, and the lowest close since 7 August 2025. This was not a risk-off session in any direction. The S&P 500 rose 0.41%, the Nasdaq Composite 1.11% and the PHLX Semiconductor index 1.61%; a fund that falls while its own equity market and the chip complex both rise is not being liquidated with anything. Nor is it gold: the gold trust fell 0.61% and the gold miners 1.41%, both less than the producers underneath the fund. And underneath it the tier split again, which is the eighth session in this account where the fund and its largest holdings disagreed: USA Rare Earth −4.27% to its lowest close since 9 July 2025 and MP −3.04% to its lowest since 29 July 2026, against Lynas +1.83% on 1.29 times average volume and Novonix +0.47%. The fund fell a third as far as the worst name it holds and rose with none of the best. That is the fifth distinct cause this section has had to record, and it is the thinnest of them: no disclosure, no policy item and no macro, on turnover below average. Then 20 October broke the run, and it is the first up-session of any size in this whole account: +1.31% to $54.06 on 716,000 shares, 0.89 times a twenty-session average of 802,845, and its best session since 21 September. The split underneath it went the other way this time, which makes it the ninth disagreement here and the first of them in the fund’s favour: REMX rose more than four of the five producers it holds — MP +0.26%, USA Rare Earth −0.30%, Perpetua −0.27%, Novonix −2.31% — and less than one, Lynas at +3.35%. Nor did the tape hand it the move: the Nasdaq Composite was flat at −0.02%, the chip index fell 0.29%, the gold trust fell 0.54% and the gold miners 0.43%. A fund that rises 1.3% while its two largest producers sit still, the chip complex falls and gold falls is being bought on its own account, which is the same finding as the selling days with the sign reversed. Year to date REMX is now 26.9% lower with a drawdown of 51.5%, against Harmonic Drive up 42.6% YTD. Still not the same exposure — and note what moved this fund across the whole run: a metals liquidation, then rare-earth-specific selling, then a security treaty, then a dollar. No humanoid order book appears anywhere in that list.
If your intent is to own the magnet bottleneck specifically, REMX is a blunt instrument for it: the cleaner expressions are a rare-earth pure play or a magnet maker, not a diversified strategic-metals basket. If your intent is to own robotics, the component names above already carry the magnet input inside their own cost base — and rising NdPr prices are a cost to them, which is part of why the co-movement between the two groups, while real at 0.39, stops well short of making one a stand-in for the other — and 23 September showed the same thing from the other end, the minerals tier down 3–7% on a dollar move while Tokyo, closed that day, reopened bid. On the magnet input itself, the position has changed, and in the direction that matters. The last like-for-like SMM figure verified here is still 4 August's $97.40/kg, and that series remains behind a paywall with only aggregators quoting it. But a second price-reporting agency has now published on the record: Benchmark Mineral Intelligence's Q3 2026 rare-earths review, dated 25 September 2026, assesses Chinese PrNd oxide at RMB729,500/t, about $108.7/kg, up 1% through September, and ex-China PrNd at $106,500/t CIF, up 12%. That is a different assessor on a different basis, so it does not retire the SMM number or get spliced onto it; it is carried as its own series. What it does retire is this page's standing claim that NdPr sits comfortably below the $110/kg DoD floor. At $108.7/kg it is within two dollars of it, which makes the floor close to the market rather than well under it, and that is a materially different fact about MP's price protection. The same review carries the number this section had been missing entirely. Heavy rare earths outside China have decoupled from Chinese domestic prices to a degree no NdPr figure captures: dysprosium oxide at $3,500/kg CIF Europe, +56% in September and 16.3 times the Chinese domestic level, terbium at $6,745/kg, +23%, and yttrium oxide CIF North America at roughly 196 times the Chinese level. That is the export-control premium priced, and it lands on exactly one name here: Lynas, the first and so far only commercial producer of separated dysprosium and terbium outside China, whose ADR fell 17.1% over the same month the ex-China price of its scarcest product rose by half. A third assessor has now printed in October, and it moves nothing: the China Rare Earth Industry Association posted its own price index at 263.6 on 9 October 2026, against the 267.0 this page carries from 4 August, with NdPr oxide quoted at ¥729.8–749.8/kg. That brackets Benchmark’s 25 September RMB729,500/t rather than contradicting it, so the Chinese domestic price has been roughly flat for a month while the index has eased about one per cent, and the $110/kg DoD floor stays close to the market rather than comfortably under it. One figure circulating alongside it is not an assessment and is not carried as one: small-volume terbium transactions reported above $12,000/kg, against Benchmark’s $6,745/kg, are trade talk relayed by a paywalled outlet with no assessor behind them. None of this turns the divergence between the two tiers into a magnet-price story, but it does mean the magnet input is no longer the quiet part of this page.
08 Method & caveats
Prices are last closes from Yahoo Finance chart data, taken as the most recent completed session in each venue, which means the as-of date usually differs by market. This run all three venues carry a new close, in the normal one-session shape rather than the weekend alignment. Tokyo reads 22 Oct 2026; Xetra and the US read 21 Oct 2026. A 07:00 UTC run on a Thursday sits after Tokyo has settled its own session at 15:30 JST, an hour and a half earlier, and after Xetra and New York have settled Wednesday, but before either Western venue opens again, so Tokyo is always one session ahead. The gap is structural to the cadence rather than a data fault, and only a weekend refresh closes it, as the 17 October run did. Xetra’s figure this run had to be recovered rather than read, and the window to recover it has already shut. Schaeffler’s 21 October daily bar is null, close and volume both, and it is null on every German line tested, Xetra, Munich, Düsseldorf and Hamburg alike, so it is the standing upstream gap described below rather than a broken ticker. The close was taken from regularMarketPrice at €5.82, stamped 17:35 CEST, five minutes after the 17:30 Xetra auction and before the 22 October session opened. By 09:06 the next morning that same field had rolled to an intraday €5.70 and the 21 October close was no longer retrievable from it at all. That is the sharpest argument this page has yet produced for running the refresh before the European open rather than after it: the figure existed for one overnight window and nothing but the quote carried it. No volume survives for the session, so Schaeffler is the one tier-one name with no volume comparison this run. Tokyo settles at 15:30 JST, not 15:00. The exchange extended its afternoon session by thirty minutes on 5 November 2024, so continuous trading stops at 15:25 and an itayose closing auction fixes the close at 15:30; every Tokyo figure here is taken at that auction, and a quote read at 15:00 is a pre-auction print rather than a close. The claim audit carries the consequence, including one instance of this page stating the old time itself. Every 1d cell on the page is still a true one-session move, but no longer over one pair of dates: Tokyo’s four tier-one cells span 21 to 22 October and the Western two span 20 to 21 October, so the 1d column still needs a date adjustment before it is read across venues. The four-week bases have split by venue, which ends the two-run single-base stretch. Tokyo’s stay pinned on 18 September for a third consecutive run, and the reason is a three-day holiday block: the exchange was closed on 21, 22 and 23 September for Respect for the Aged Day, a Citizens’ Holiday and the Autumnal Equinox, so a four-week window ending on 22 September cannot reach any session later than 18 September. The Western names’ bases rolled to 21 September. The tier gap this run, 8.3 points to 7.4, therefore carries calendar where last run carried none: held on a common 18 September base it would read 6.5, so price narrowed it 1.8 points and the base roll widened it 0.9. The largest single effect sits in MP, whose four-week base rolled $47.26 to $50.00 and took its 1m cell from −16.4% to −21.0% on a session the stock rose 1.15%. One column does not follow that rule, and it is worth stating plainly: fiftyTwoWeekHigh is anchored on the request date rather than on the venue’s last completed close, so the drawdown lookback window can roll on a day the price behind it did not. That is what happened again on 16 October, when the window moved onto the 17 October 2025 highs and took MP’s from $91.99 to $84.15 and USA Rare Earth’s from $34.03 to $32.80, and it is the mechanism this page wrongly denied for weeks. It rolled again on 17 October and that time barely registered: only MP’s high moved, $84.15 to $83.99, two tenths of a per cent, with the other ten unchanged. It did not roll at all on the 19 October run, when all eleven 52-week highs held. On the 21 October run two moved again, and they were the same two: MP’s high rolled $80.89 to $76.80, improving its cell from 51.9% to 49.2% and taking it under fifty per cent for the first time in this account, on a session the stock rose a quarter of a point; and Novonix’s rolled $21.50 to $20.70, which left its cell all but still at 89.8% because the price fell 2.31% into it. MP’s is the largest step this column has taken in a single run, and the decomposition is stark: held at today’s price on the old $80.89 high the cell would read 51.7%, so of the 2.7 points, price contributed 0.2 and the 20 October 2025 high dropping out of the window contributed 2.5. The other nine were unchanged. On this run not one of the eleven moved, so every change in the drawdown column is price, and in Tokyo there was a great deal of it: Harmonic Drive’s cell widened 5.6 points in a single session to 46.7%, Nabtesco’s 3.2 to 31.9%, Yaskawa’s 2.1 to 49.0% and Nidec’s 1.6 to 41.7%, while Ambarella’s narrowed 2.9 points to 31.7% and is now the tightest drawdown on the page. The larger base movement this run is again in the one-year column, and it is the exact mirror of the last one. Tokyo’s bases rolled from 21 onto 22 October 2025 and the Western names’ from 20 onto 21 October 2025, which takes 20 October 2025, the session the market traded MOFCOM’s extraterritorial controls and the Western producers re-rated on scarcity, back out of the window. Last run that session rolling in cost four minerals cells between four and six points; rolling out, it hands back more than it took. Perpetua’s 1y goes −31.4% to −21.0% on a base rolled $27.27 to $23.71, and 10.3 of those 10.4 points is the base with a tenth of a point of price; Lynas’s −42.0% to −32.9% on $13.84 to $11.99, 9.0 points of it base; USA Rare Earth’s −68.2% to −61.7% on $31.59 to $26.73, 5.9 of base; MP’s −52.9% to −46.7% on $82.90 to $74.11, 5.7 of base; and Novonix’s −90.6% to −89.3% on $22.50 to $20.00. Tokyo’s step ran the other way and was swamped by price: Harmonic Drive’s 1y falls from +65.6% to +49.1% on a base rolled ¥3,255 to ¥3,270, of which only 0.7 of the 16.5 points is the base and 15.8 is the session itself; Nabtesco’s +17.1% to +8.3% splits 3.6 to base and 5.1 to price, and Yaskawa’s −3.4% to −11.6% splits 4.2 and 3.9. The rule is unchanged and this run states it in both directions at once: read the 1y column with the base in hand.
Both documented feed traps appeared in the same ticker on consecutive runs, pulling opposite ways, which is the clearest demonstration of why the rule is split in two. Last run the German daily array carried a 12 October bar at €5.97 stamped 09:04:22 CEST on 57,043 shares, an intraday quote masquerading as a close, and it was discarded because the venue was open. This run that same session has become a hole: the 12 October daily bar is null, and the close had to be recovered rather than read. Two independent routes agree on €5.95: regularMarketPrice carrying it stamped after the 17:30 auction, captured in the narrow window before that field rolled forward to a live 13 October print of €5.93 on 5,823 shares, and previousClose returning the same €5.95 on a five-minute request made after Xetra reopened. The continuous five-minute series is the trap inside the trap: it stops at 17:25 on €5.89, and taking that last tick would have published a flat session rather than the +1.19% one that happened. So the rule holds in both halves. If the venue is open, the last bar is discarded regardless of what it says; and when a bar is missing, prefer the auction print over the last continuous tick. That was the twenty-fifth hole in this series. The gap has recurred on both sessions since, at 13 and 14 October, while 12 and 13 October have each backfilled at exactly the figure published here, €5.95 and €5.85. The 14 October bar backfilled at exactly €5.94 and the 15 October bar has now backfilled at exactly €5.75, both the figures published here from three routes, so the reconstruction that stood open last run is confirmed. 16 October is null in its place, and this one is a genuine hole rather than the ordinary open-venue case, because Xetra had been closed for thirteen hours when this page compiled. 16 October has since backfilled at exactly €5.83, the figure published here from two routes, and a thirtieth hole opened in its place on 20 October. So the standing tally is thirty holes, thirty reconstructions, twenty-nine confirmed to the cent, with 20 October the only one open, closed here at €5.89 on two routes: regularMarketPrice stamped 17:35:05 CEST against the 17:30 close, and previousClose returning the same €5.89 on a request made after Xetra had reopened. This one is a deeper hole than its predecessors. The 16 October bar had made the shape of these gaps unusually clear, its volume field populated at 877,225 shares while its close was null, so the session was in the feed and only the price was missing; the 20 October bar drops the volume too, which is a different and worse failure, and the twenty-session comparison had to be made against the quote-level volume field instead. The trap inside the trap also pulled the opposite way from last time: the continuous five-minute series stops at 17:25 on €5.94, so taking the last tick rather than the auction would have published a +2.95% session instead of the +2.08% one that happened. That session also produced this series’ first corroboration from outside the vendor altogether: a German market report stamped 15 October 13:20 CEST quotes the preceding “Xetra close of EUR 5.94 on October 14, 2026”, matching the reconstruction rather than the feed that lost it. One route was weaker than last run’s, and it should be recorded as such: the 17:30 five-minute bar carried €5.75 on zero volume, so it is the continuous close rather than the auction print, where 14 October had shown the auction itself on 1,221,280 shares. The largest moves on the page this run are in a column again, and this is the biggest base roll the series has recorded. Every twelve-month base stepped from 9 onto 10 October 2025, one session on from the one it occupied last run, and that single day does something none of the others did: it split the page. 9 October 2025 is the session MOFCOM published Announcements 61 and 62, the extraterritorial rare-earth export controls; 10 October 2025 is the day the market traded them, and the Western producers re-rated on scarcity while the broad complex sold off on the tariff threat that followed. So the bases moved in opposite directions and the twelve-month cells followed them. Novonix’s base rose 35.8% in one step, from $21.65 to $29.40, which is the largest single base movement this page has seen in any column, leaving its cell at −91.8%. Compass Diversified’s rose 13.6%, from $7.35 to $8.35, and its cell fell 18.5 points, from +47.6% to +29.1%, on a −0.65% session. MP’s rose 8.4%, from $72.29 to $78.34, for an 8.5-point fall to −44.7%, and USA Rare Earth’s 5.0% for a 4.2-point fall to −63.8%. Ambarella went the other way and furthest: its base fell 13.5%, from $84.28 to $72.88, so a 3.89% loss left its cell 8.9 points better, from −20.0% to −11.1%. The instructive one is REMX, whose own base fell 7.1%, from $74.78 to $69.47: on the session that priced China restricting rare-earth exports, the rare-earth ETF fell while three of the five producers it is meant to proxy rose, which is this page’s third finding stated by a single day of data. The four-week bases, by contrast, did not move at all, and this note’s prediction that 13 October would be the first chance for the venue alignment to break was wrong. All seven are still 11 September, for a fourth consecutive run, because 12 and 13 September were a weekend: a 30-day lookback from 12 October and from 13 October both land back on the same Friday close. The alignment the previous run called a coincidence of the calendar turns out to be a fairly durable one, and it will hold until a lookback clears a weekday. Read every column against its own base date, and treat a large move in one on a small move in the price as a date problem until shown otherwise. A venue date is still not a page date; read each column against its own. One trap, recorded on 19 Aug 2026: Yahoo’s daily bar array can lag a venue’s latest close, and that afternoon it still ended at the 18 August Tokyo session while the quote metadata already carried the 19 August close. Taking the bars at face value would have shown Harmonic Drive at ¥5,960 rather than ¥5,670, understating the fall by about five percentage points. The German feed keeps swallowing sessions, and the running tally is the argument for the method rather than against it. The deepest case in the series stays closed and confirmed: 24 September, the largest session of the series at −6.2% and the only one closed on four independent routes rather than two, backfilled at exactly €6.54, the figure the auction bar, previousClose, chartPreviousClose and an independent German quote service all carried at the time. That took the tally to sixteen holes, sixteen reconstructions, sixteen confirmed to the cent through the end of that week. It has not stayed there for a single run since. The nineteenth hole, at 1 October, has now healed, and it is the strongest confirmation in the series because it was the one closed on three independent routes rather than two. That bar now reads exactly €6.23, which is what all three carried when it was absent: regularMarketPrice stamped 17:37:57 CEST against a 17:30 close, previousClose on a five-minute request made after Xetra reopened, and a five-minute bar stamped exactly 17:30:00, the auction itself rather than a proxy for it. Three routes, one figure, and the vendor has now agreed with all three. The twentieth hole, at 5 October, was the first in the series that could not be closed on a second daily route, and it has now healed on the vendor's own bar at exactly €6.24 — the figure regularMarketPrice carried stamped 17:39:37 CEST against a 17:30 close, in the single narrow window it was available, since on the next request the same field had already rolled to a live print. That capture window is the whole of the method here, and the single-route reading has been vindicated rather than merely left standing. A twenty-first hole opened immediately behind it, at 6 October, and that one closed on two routes: regularMarketPrice at €6.36 stamped 17:35:22 CEST, and a five-minute bar stamped exactly 17:30:00 carrying the same €6.36, which is the auction itself rather than a proxy for it. The Frankfurt line carries the identical hole for a second consecutive session, which again confirms the gap is one upstream failure rather than a broken Xetra ticker, and again removes the second German daily route in the same stroke. So the tally reads twenty-one holes, twenty-one reconstructions, twenty-one confirmed to the cent, and that sentence is only true as of this run: when it was first published the 6 October bar was still null, so the twenty-first case was reconstructed rather than confirmed, and this note said “none open” while the entry above it correctly reported the bar as missing. The vendor has since backfilled 6 October at exactly €6.36, which is what both routes carried, so the claim has caught up with itself and nothing published ever depended on the difference. The twenty-fourth hole, at 9 October, has now healed as well, and it was one of the better-closed cases in the series rather than a single-route reading. The vendor's bar now reads exactly €5.88 on 932,268 shares, which is what both routes carried when it was absent: regularMarketPrice stamped 17:39:43 CEST against a 17:30 close, and a five-minute bar stamped 17:35:00 at the same €5.88 on 435,144 shares, against 2,745 to 18,539 in each of the six bars before it. That took the tally to twenty-four holes, twenty-four reconstructions, twenty-four confirmed to the cent, and none open, on the session that took Schaeffler to its lowest close since 2 October 2025. Three further holes have opened since, at 12, 13 and 14 October, and the first two have already backfilled at exactly the figures published here, so the standing tally is twenty-seven holes, twenty-seven reconstructions, twenty-six confirmed to the cent with the newest open. The procedure has been contradicted on none of the twenty-six closed cases, which is as settled as a workaround gets. It stays in place, because the gap has recurred often enough that its absence for one session would not be evidence it had stopped. The rule stands with its clause: when a German bar is missing, prefer the auction print, and where the intraday series stops short of the auction, previousClose on a request made after the venue reopens is the auction, not the last tick before it. The US feed's own hole has healed too. The 22 September daily bar was missing for four days for Ambarella, Lynas's ADR, Perpetua, Novonix, REMX, Compass Diversified and the iShares gold trust, while MP, USA Rare Earth and the SPDR gold trust carried it throughout — the same split, and mostly the same names, as the 28 August attrition recorded below. All seven were closed from previousClose and all seven have now returned, matching the reconstruction on every one: Ambarella at $67.06, REMX at $70.97, Perpetua at $24.76, Lynas at $10.33, Novonix at $2.85. That last one is the figure the claim audit's newest row turns on, and it is worth noting that the page's table had it right on the day and only the prose did not. The same gap hit an earlier run from the opposite direction, when the daily bars for every US name ended at 2 September while the metadata already carried the completed 3 September close. The metadata timestamp, converted to exchange-local time, is what decides whether a session is complete. Returns are computed on the local primary line, so no currency translation is embedded. This is a dated snapshot — the page cannot fetch live quotes and does not pretend to. It is refreshed daily at 07:00 UTC, and the compiled date in the masthead moves every time any content on this page changes, so that stamp is always the age of what you are reading.
28 Aug 2026: the corporate-action trap. Novonix changed its ADS ratio on 27 August, effectively a one-for-ten reverse split of the US line. Yahoo adjusted its quote metadata immediately — the 52-week high became $38.60 — but had not yet restated the daily close history, which still ran to $0.435. Read together, the two would have put Novonix 98.9% below its 52-week high, an invented 10× drawdown. Yahoo restated the history in stages, and for a fortnight put the seam in the wrong place: the 10× step sat between 21 and 24 August, three sessions before the change took effect on 27 August, so 24–26 August were carried at post-change prices they never traded at. That restatement has now completed. The whole series is on the post-change ratio, the misplaced step is gone, and a scan of two years of daily bars finds no jump larger than the two genuine ones from November 2024 and October 2025. The figures on this page were computed throughout with both sides on the post-change ratio, scaling closes at the seam the data actually contained rather than the one the corporate action implied — and they did not move when the vendor caught up, which is the only real test such a patch gets. Taken raw at the time, the same mismatch would have reported Novonix up 618% in a month, the earlier error wearing the opposite sign. Any screen mixing adjusted metadata with unadjusted bars across a corporate action produces the same class of error, silently.
A hole in the German feed, since filled. Yahoo's daily bars carried no 28 August close for any German single stock — Schaeffler, SAP, BMW and Continental all return null for that session, as do the Frankfurt, Munich, Düsseldorf and Hamburg lines, so it is one upstream gap rather than several — while the DAX index itself printed 26,570.0 that day. Taking the bars at face value would have held Schaeffler at its 27 August close of €7.35 and reported a 2.4% rise as a flat session. The €7.53 on this page is reconstructed from the intraday series for 28 August and cross-checked against an independent price history that agrees to the cent and matches Yahoo's own daily bars for 24–27 August exactly. A missing bar is not the same as an unchanged price. Yahoo backfilled the session later the same day at €7.53, matching the reconstruction to the cent.
The German hole is now a standing feature — and it heals itself, which is the useful part. It has recurred on eleven sessions since late August, and it is not Schaeffler-specific: SAP and Siemens return null for exactly the same days, so it is one upstream gap rather than a broken ticker. What has changed is that the earlier gaps have since backfilled, which turns them into a test of the patch. Yahoo now carries daily bars for 28 August and 1–4 September, and every one of them matches the figure this page reconstructed at the time to the cent: €7.53, €7.22, €7.02, €7.30 and €7.43. Five reconstructions, five exact confirmations, using two different routes — the neighbouring session's quote metadata, and the intraday series. The 7 and 8 September holes needed a third route, because both were still missing while Xetra had already opened on 9 September, so the metadata carried a live quote rather than either settlement and the standard 5-minute series stops at 17:25, before the closing auction. Requesting the 5-minute series with pre- and post-market bars included returns the auction print itself, stamped 17:35 CEST: €7.34 on 7 September and €7.43 on 8 September, the latter on 229,007 shares against a few thousand in each of the preceding bars — the volume signature of a closing auction rather than a stray tick. The 8 September bar has since backfilled at exactly €7.43, a sixth reconstruction confirmed; the 7 September bar has still not returned at all. The 9 September close, €7.40, came back a fourth way and the simplest one yet: the daily request, made before Xetra's 10 September auction, still carried 9 September's settlement in regularMarketPrice stamped 17:39 CEST, and the 5-minute request made at the same moment returned the same €7.40 in its previousClose field. Two independent fields, one figure — and the vendor has since backfilled that session at exactly €7.40, a seventh reconstruction confirmed. Taken from the bars alone the page would have carried Schaeffler at €7.43 and called a live session unchanged. 11 September behaved the same way: €7.19 in regularMarketPrice, stamped 17:35:12 CEST against a 17:30 close, while the daily bars stopped at 10 September — and that session has since backfilled at exactly €7.19, an eighth reconstruction confirmed. 14 September has since backfilled at exactly €6.95, a ninth reconstruction confirmed, and 15 September at exactly €6.78, a tenth — that one having been closed by two independent routes on the day it opened, the metadata auction stamp and the 17:35 five-minute bar on 575,656 shares, rather than after the fact. 16 September took a different second route again: €6.86 in regularMarketPrice stamped 17:38:13 CEST, with the five-minute series stopping at 17:25 and never printing the auction, so the corroboration came from previousClose on a request made after Xetra reopened — and that session has since backfilled at exactly €6.86, an eleventh reconstruction confirmed. So has 7 September, at exactly €7.34, a twelfth — the hole that stayed open for a fortnight, and the only one that needed the extended-hours request to close at all. 21 September closed the quickest of that group: the five-minute series printed the 17:30 auction at €6.99 on the day, matching regularMarketPrice at 17:35:12 CEST, and it has since backfilled at exactly €6.99, a thirteenth reconstruction confirmed. 22 September followed at exactly €7.13, a fourteenth — the most instructive of them, because the intraday series stopped at 17:25 showing €7.16 and the page took previousClose over it; had it taken the continuous print it would have published a 2.4% session rather than a 2.0% one. 23 September has now backfilled at exactly €6.97, a fifteenth. Sixteen holes, sixteen reconstructions, fifteen of them confirmed to the cent; only 24 September, the largest session of the lot at −6.2%, is outstanding — and it is the one closed on four routes rather than two. That is settled enough to be a procedure rather than a surprise — when a German bar is missing, take the auction print from the metadata timestamp and check it against a second field before believing it.
Bars that come and go. Novonix's daily history around its 27 August ADS ratio change is not stable between requests: 28 August read $3.51 at one run and had disappeared again, along with 26 and 27 August, by the next. Left alone, that would have moved the name's correlation and shown a four-session fall as a one-day move. The 28 August close is pinned to the verified figure so the series does not wobble with the feed. By 3 September the instability had spread well beyond Novonix: the whole 28 August US session had gone missing for Ambarella, REMX, Perpetua, Lynas and Compass Diversified while surviving for MP and USA Rare Earth — per-ticker attrition, not one dropped day. That has since healed. All five carry 28 August again, and Novonix's bar for it has returned at exactly $3.51, the figure this page pinned when it kept vanishing. The weekly series takes each ISO week's last available close, which absorbs this without shifting a week; where it removes a Friday, that week ends on the Thursday instead. Where a vendor's history changes underneath you, the fix is to record what you verified, not to re-derive it each run.
4 Sep 2026: a continuous futures series that is not continuous. This page has priced gold off GC=F, Yahoo's front-month gold contract, to sanity-check Perpetua's moves. That series silently rolled: GC=F now resolves to Gold Dec 26, and its own bar array still carries $4,366.3 for 2 September — the expiring contract — while the December contract closed that session at $4,414.6. Differencing across the roll would have reported a +4.0% gold session on 3 September against +1.9% for the bullion ETFs, which agree with each other to four hundredths of a point (SPDR Gold Shares +1.85%, iShares Gold Trust +1.88%). A futures move measured across a contract change is a spread, not a return. Gold figures on this page are now taken from the spot-tracking ETFs, which also settle on the same 16:00 ET clock as Perpetua and the miners they are being compared with — the earlier $4,366/oz reference has been dropped for the same reason.
Correlations use 103 completed weekly returns bucketed by ISO calendar week so Tokyo, Xetra and New York lines align despite different trading calendars. Weekly rather than daily specifically to avoid the time-zone lag that inflates or deflates cross-market daily correlation. The window runs to the completed week ended 16 Oct 2026 (ISO week 42); it was extended from 53 weeks on 14 Aug 2026, which shifted several tier-two figures down by 0.10–0.20 without changing any conclusion. The basket is equal-weighted and rebalanced weekly — the basket return for a week is the mean of the six tier-one names' returns for that week — and stating that is not pedantry, because the same six names held without rebalancing give 0.3993 against the rebalanced 0.3877 on the current window. This page published 0.41 through 16 Sep 2026 without naming the construction, which made the figure impossible to rebuild from this note; the claim audit carries the full account. Every figure is cross-checked against a second bucketing method before publication, and on this run the two agreed to four decimal places on every series. The window did not move on the 20 October run, and three cells carried on it were still slightly wrong. ISO week 43 opened on Monday 19 October and does not close until Sunday 25 October, so the sample holds at 104 aligned weekly closes and 103 returns, week 2024-43 to week 2026-42, exactly where the 17 October run set it, and basket-to-REMX holds at 0.3885 on both methods. On an unchanged window a cell cannot move, so a cell that recomputes differently was carried wrong rather than overtaken. Three did: Lynas to REMX at 0.6471 against a published 0.64, Nidec to the basket at 0.6229 against 0.61, and Harmonic Drive to the basket at 0.8152 against 0.81. Each rounds a point higher than the figure on the page, each was published on a window identical to this one, and all three are corrected here to 0.65, 0.62 and 0.82 in both bar charts and in the tier-two table. The claim audit carries it. None of them changes a reading: the robotics names still run 0.18 to 0.46 against REMX and 0.55 to 0.82 against the basket, and the minerals names 0.37 to 0.65. The account that follows belongs to the run on which the window last advanced, and it did so on a judgement rather than a rule, so the judgement is kept on the record rather than left for a reader to infer from a moved number. ISO week 42 runs to Sunday 18 October, so on a weekday run it would still be in progress and excluded, which is what the last three runs did. This run is a Saturday: every venue has settled Friday 16 October and none reopens before Monday, so week 42’s last close is final even though its calendar week is not over. The week is therefore counted as completed, the sample now ends at the week ended 16 October. The count published with it, 104 returns against 103 on the preceding runs, was wrong: a two-year request rolls forward at both ends, so the week gained at the end is paid for by one dropped at the start, and the aligned sample held at 104 closes and 103 returns. The judgement about which week ends the sample stands; only the count was overstated, and it is corrected throughout on the 19 October run. That is a shift from “the calendar week has ended” to “the week’s trading has ended”; it can only ever differ on a weekend run, and it is the kind of definitional change this page has an audit row about making silently. Week 42 was heavy on both sides and they fell together: the basket −8.31% against REMX −11.05%, with Harmonic Drive −13.99%, Nidec −10.45%, Ambarella −8.50%, Yaskawa −8.47% and Nabtesco −7.64%, against Schaeffler nearly flat at −0.85%, and in the minerals tier USA Rare Earth −16.08%, Novonix −13.93%, MP −12.94%, Lynas −12.60% and Perpetua −6.93%. Two tiers falling hard in the same week sits close to the existing relationship rather than against it, so the headline figure rose slightly and no conclusion in section 04 moves: basket-to-REMX goes from 0.3687 to 0.3877. Both bucketing methods were recomputed from freshly fetched data and agreed on every series, 0.388 each on the headline figure. The one name the new week moved against the grain is Schaeffler. It was the only line in the basket not to fall materially in a week the basket fell eight per cent, and that pulled its correlation to the basket down, from 0.6070 to 0.6016 — the exact mirror of the mechanism recorded below, where being the worst line in a falling week had pushed the same figure up. The account that follows is the record of the last run in which the window actually advanced. Week 41 closed on Sunday 11 October, so the end advanced from the week ended 2 October to the week ended 9 October, and the whole of a busy week entered the sample in one step: TDK's four-session run and the three losses that followed it, Monday's Tokyo rally, Thursday's de-rating of the gearbox layer, Friday's partial reversal, Yaskawa's first-half filing, two sessions of rare-earth-specific selling and Friday's seven-point move in Ambarella. The headline figure moved seven ten-thousandths, from 0.3681 to 0.3687. The largest change anywhere on the page is Schaeffler, 0.5997 to 0.6070 against the basket and 0.4564 to 0.4645 against REMX; Lynas goes to 0.6333, MP to 0.5524, Perpetua to 0.5248, USA Rare Earth to 0.3678 and Novonix to 0.3603. Both bucketing methods were recomputed from scratch on fresh data and agreed to four decimal places on every one of the fifteen series, basket included, which is the stricter of the two outcomes this check produces. One prediction in this note was right in its number and wrong in its reason, and the correction matters more than the figure. It said the week would “substantially cancel”. It did not cancel: week 41 was a broadly down week on both sides, the basket 3.27% lower and REMX 3.54% lower, with nine of the eleven charted names falling. The figure barely moved because the two sides fell by nearly the same amount, which places the new observation almost on the existing line and so confirms the relationship instead of altering it. Equal-sized falls on both sides and mutual cancellation inside one side are different mechanisms that happen to produce the same small change, and only one of them was what occurred. The mechanical rules all held: a window ending at the last completed ISO week cannot move inside an open one, so neither a Friday close nor a Saturday refresh advances the end, which advances on the first refresh after a Sunday — the 12 October run was that refresh, and no run inside the same week can repeat it, which is why this run still holds that window. The start, separately, rolls forward on its own, which was the lesson of 16 Sep 2026 and is visible here: the sample start stepped from ISO week 2024-41 to 2024-42 as the end advanced, holding the length at 103. And a standing claim in this note has to be withdrawn, because testing it is what this check is for. The note has said that a two-year daily request yields 103 weekly returns rather than 104, that a true 104-return window needs a three-year request, and that the figures here are built from one. The first two are right; the third is not. Recomputing from a three-year request at a genuine 104 returns gives, on the window as it now stands, basket-to-REMX 0.3710, Lynas 0.6336, Schaeffler-to-basket 0.6022 and Harmonic Drive-to-REMX 0.1629, none of which is what this page publishes; the same test on the window ending 2 October gave 0.3670, 0.6322, 0.5991 and 0.1581. Every figure here reproduces exactly on 103 returns and on nothing else, so 103 is what the window is, and the counts in the charts and captions have been corrected to say so. The difference is about a thousandth and changes no conclusion, which is precisely why it survived: a discrepancy too small to notice in the output is only ever caught by rebuilding the claim rather than re-reading it. The four-week comparison between the tiers uses a different construction again, and it is now stated: the basket's 1m figure is the mean of the six tier-one names' own 1m cells — equal-weighted, not rebalanced, each name on its own venue's four-week (28-day) lookback — so a reader can rebuild it from the table above. This page previously quoted 5.2% for that figure and it cannot be reproduced on any construction tried; the claim audit carries that too. Correlation measures co-movement, not causation — treat 0.37 as "these travel together loosely," not as a precise constant.
A correction worth repeating, and a bigger one below it. Schaeffler's correlation to the robotics basket was originally computed on the SFFLY ADR and came out at 0.16; on the current window the same thin line gives 0.29. On the liquid Xetra line it is 0.60. Thin ADRs do not merely lag — they can invert a conclusion. Every correlation on this page now uses the primary listing.
18 Aug 2026 correction: the alignment bug. Every correlation figure this page had published through 17 Aug 2026 was computed from Yahoo's native weekly bars (interval=1wk). Those bars anchor Tokyo and Xetra weeks roughly one calendar week off from New York's — confirmed directly by comparing native bars against daily closes re-bucketed into true ISO weeks, which showed a consistent one-week offset for every Tokyo and Xetra ticker tested but not for US tickers. Because the tier-one basket is four-sixths Japanese names, this silently corrupted every cross-region figure: the robotics basket's correlation to REMX read 0.19 on the misaligned data and about 0.40 once daily closes were re-bucketed by calendar week (0.39 on the current window, weekly-rebalanced) — confirmed by two independent methods (ISO-week bucketing and Friday-anchored resampling) that agree to three decimal places. Correlations computed entirely within one region (e.g. TDK, a Tokyo name, against the mostly-Tokyo basket) barely moved, which is exactly what a same-direction shift on both sides of a comparison would produce, and is why the bug went unnoticed until a cross-region figure was checked against an independent method. Every correlation on this page, in every table, chart and paragraph, is now computed from daily closes re-bucketed into ISO calendar weeks.
Use the primary listings. The ADRs quoted in the source post are thin: HSYDF traded 200 shares in a session, YASKY 513, NCTKY 1,350. NJDCY has not printed since February. Prices on those lines can sit stale for days and will not reflect Tokyo. Where an ADR is the only access route, expect a spread.
Not investment advice. This is a research snapshot assembled from public filings, company releases and press reporting, built to check a social-media thesis rather than to endorse it.