The useful check isn't what to copy. Multiply your biggest holding's weight by a plausible fall: 25% of a portfolio falling 30% takes 7.5% off the whole balance before other holdings move. Check whether other positions could fall at the same time, and whether you could keep contributing or meet planned spending without selling.
The disclosed book became more concentrated going into the October-November slide in chip and storage shares. Held at fixed share counts, the September shares fell 16.1% from 30 September to 18 November, while the MSCI World price index in US dollars fell 4.3%. Natera's 30% fall accounted for 9.5 points of that loss and Seagate's 44% fall for 3.4. That is a hypothetical: the family office may have traded since, and its shorts, futures and bonds are outside the filing.
What is in the fund?
The 13F shows 65 lines in 61 securities at 30 September 2026, all long shares or calls. The ten largest securities make up 69.9% of reported value, up from 53.2% in June. Natera is 29.2%, Seagate 9.1% and Amazon 6.9% including calls, Fox class A shares 4.3%, Insmed 4.2% including calls, YPF 4.0%, Unity Software 3.6% and Sea 3.5%.
A useful single number for concentration is the effective number of positions, which counts how many equal-sized holdings would give the same concentration. It fell from 20.4 in June to 9.1 in September. Calls, counted at the value of the shares they cover as the 13F requires, fell from 16.4% to 7.2% of reported value.
30 June to 30 September 2026, share of reported value: Natera 16.6% to 29.2%; data storage 3.8% to 9.9%; other chips 13.0% to 3.1%; broad-market ETFs 6.8% to 0.3%; gold miners 0.2% to 3.1%; copper, steel and aluminium 1.4% to zero. Source: 13F-HR information table, 30 September 2026 and 30 June 2026; every line from December 2025 to September 2026 (CSV). The themes are our grouping of the filing's lines.
What changed since 30 June
Natera's jump was a price move, not a purchase. Duquesne held 3,160,666 shares, 0.8% fewer than in June, while the price rose from $271.45 to $413.31. The other big shifts were trades.
- Fewer chip holdings. TSMC fell from 589,680 to 32,700 shares, STMicroelectronics was nearly halved, and AMD, Lam Research, Entegris, Rambus and Arm are absent from the September filing.
- Storage in. Seagate rose from 122,000 to 341,790 shares, with new calls on 98,000 more, and Western Digital is new; Sandisk, held in June, is gone.
- Index calls absent. Calls on the equal-weight S&P 500, Russell 2000, S&P 500 and MSCI Brazil funds covered 9.3% of June table value; none were reported in September.
- Different metals exposure. Newmont, Barrick and Agnico Eagle are new, while Southern Copper, Cleveland-Cliffs and Alcoa are absent.
- Other exits. Alphabet, United Airlines, Delta and Teva, which Druckenmiller had used in January as an example of Duquesne's process.
Unity Software grew from 738,785 to 3,956,385 shares, and Palantir and SpaceX are small new positions. In all, 16 securities are new and 45 have gone; the new ones make up 9.2% of September value.
What the filing cannot show
The SEC's 13F guidance excludes short positions and shares that trade only on non-US exchanges. The form covers securities on the SEC's official list, mainly US-listed shares and options, so futures, currencies and government bonds are outside it. For a macro investor, that leaves out most of the levers.
In January 2026 Druckenmiller described big positions in Japan and Korea, a bearish view of the dollar, copper futures and a short bond position, none of which a 13F would show. Our earlier study of his strategy covers that gap and eight quarters of his filings in detail.
Why does Druckenmiller use this structure?
He describes his method as a few large ideas, changed quickly when the evidence turns. In a Morgan Stanley interview recorded on 30 January 2026, he explained the approach eight months before the September snapshot. The interview does not explain the individual September trades.
In January he said AI had become less central to his portfolio. He described AI valuations from mid-2025 as reminiscent of the late 1990s, while acknowledging that the portfolio still had some AI exposure.
He described moving into biotech because it might lead the market and make productive use of AI. He said he relied on the team's expertise in the science. In September, Natera and other healthcare and biotech names together made up 41.5% of the disclosed value; that does not establish why any individual position was held.
He described his gold exposure in January as mainly a geopolitical trade. The September filing shows gold miners, which are businesses, and we found no statement that they serve the same purpose.
He credits George Soros with teaching him to pay more attention to how much he gains or loses on a trade. He also described reversing an intended three-year trade after five days. A quarter-end snapshot filed 47 days later shows the positions but not those exits.
No return series exists for the money Druckenmiller runs today, so no matched geometric return, Sharpe ratio or Sortino ratio against MSCI World can be calculated. Duquesne Family Office publishes no returns. The only long-run figure is for Duquesne Capital, which he closed in 2010: Bloomberg, as reported by Financial Planning on 18 August 2010, said its returns "had averaged 30 percent annually since 1986".
That figure has no annual path, no stated fee basis and no volatility, so it cannot be compounded, risk-adjusted or set against an index over matched years. Changes in the 13F value column mix price moves, purchases, sales and the value of shares underlying calls. They are not returns. The comparison that can be made honestly is with the disclosed shares themselves, in the next section, and it describes the shares, not the family office.
Copio's reading
Is it optimal for geometric growth?
The disclosed book is a concentrated growth bet whose sizing would be justified only by a large expected edge in Natera; optimality cannot be shown from the filing. The objective we test is expected log wealth, meaning growth that compounds through gains and losses, over a decade or more. The family office was announced in 2010 to oversee his own wealth, so outside redemptions should not constrain him; the comparison set is the same capital in a world index fund, or the same book with a smaller largest position.
First, is the concentration paid for? Natera's daily returns had an annualised volatility of 48.2% over the year to September. In an illustrative model with only Natera and cash, a 29.2% weight corresponds to about 6.8% expected excess return using the full-Kelly approximation, or 13.6% if using half the model's weight. That is not a forecast or an estimate of Duquesne's actual position size. Natera had a 0.36 correlation with the other disclosed shares, and the filing does not give the full portfolio inputs needed to size the real position.
Volatility at that level also has a cost of its own: roughly, a stock's compound growth trails its average yearly return by about half its variance, here about 11.6 points a year. The stock rose 52.3% in the third quarter and fell 30.0% in the seven weeks after it, which shows the swing without telling us the edge. Our Kelly criterion guide shows the same arithmetic for a whole equity allocation.
Second, did the changes buy protection that pays for itself? Closing the index calls removed leveraged market exposure, but in this sell-off the indexes barely moved: the S&P 500 fund fell 4.3%, the Russell 2000 fund 2.8%, and the equal-weight fund rose 3.6%. The loss came from stock-specific positions, not from the market exposure that was cut. The premiums paid for the calls are not in the filing, so their net cost cannot be measured.
Third, do fees, financing and cash flows change the result? Fees and financing costs are not public, including what was paid for derivatives or borrowing. The filing does not establish that the expected returns compensated for concentration. Whole-portfolio returns and risk exposures would let us judge whether the positions worked together through losses; a single window cannot do that.
Is it crash aware and protected?
The disclosed shares took a large loss in the hypothetical; the filing cannot establish whether the full portfolio was protected. The September 13F lists no puts. Held at fixed share counts from 30 September, the shares fell 16.1% by 18 November, against 4.3% for the MSCI World price index in dollars. Neither figure includes dividends.
Hypothetical, 30 September to 18 November 2026: September shares −16.1%; June shares −15.3%; MSCI World price index −4.3%. Source: 13F-HR for 30 September 2026, Yahoo Finance daily closes and MSCI World price index in USD; daily path (CSV), price inputs (CSV).
The September share counts produced a larger loss than the June counts in this window. That comparison does not establish whether the new holdings were riskier in general. Seagate fell 44.1% and Western Digital 36.6%, against 20.5% for TSMC, so the storage trade lost more than the chip position it replaced.
Natera alone took 9.5 points off the September book and Seagate 3.4; Bitdeer fell 62.9%. CDW (+31.9%), BBB Foods (+29.5%), Unity (+19.6%) and Amazon (+7.5%) rose. Recovering from the 18 November level needs a 19.1% gain; our recovery maths article explains why losses need larger gains to repair.
None of this is the family office's result. Druckenmiller's stated defence is the exit, and the filing cannot show whether he held these positions through November. In January he also described a short bond position to manage the risk of higher yields. Such a position can gain when bond prices fall, but it is not a general equity-crash hedge. An exit rule is not a hedge: it works only if he sells before the loss, and the 13F records only quarter-ends.
Is it diversified without sacrificing growth?
The largest positions did not usually move together, but their sizes still mattered: two holdings drove most of the hypothetical loss. A correlation of 1 means two daily returns move in lockstep and 0 means no linear relation. Over the year to 30 September 2026, the ten largest securities had pairwise correlations from -0.15 to 0.26, averaging 0.11.
30 September 2025 to 30 September 2026, 251 daily returns: pairs -0.15 to 0.26, average 0.11; Natera with Seagate 0.08; Amazon with the Brazil ETF 0.26. Source: Yahoo Finance daily closes; correlation matrices for both windows (CSV). These are holdings correlations, not a measure of the whole family office.
In the 35 trading days after 30 September the average stayed at 0.12, so the holdings did not start falling together. The book still fell nearly four times as far as the index, because Natera and Seagate fell for their own reasons and together were 38.3% of reported value. Low correlation reduces the chance that everything falls at once; it does not limit what one oversized holding can do.
Swapping TSMC for storage did not diversify the chip risk. Seagate's correlation with TSMC was 0.52 over the year and 0.69 after September, so the new position stayed close to the same AI hardware risk.
The gold miners did more: their daily correlation with the other disclosed shares fell from 0.51 over the year to 0.25 after September, and they were 3.1% up on 6 November, when storage was down 49.2%. They still ended 11.9% lower, so they softened the path without offsetting the loss. Our study of gold in crashes shows why miners and the metal behave differently.
Copio does not grade this portfolio in its checker, because a book of single stocks and options with undisclosed shorts, futures and bonds cannot be represented by an ETF replica.
What a household can take from it
A 13F is a quarter-end snapshot filed weeks later. It can show listed puts but leaves out many other ways a manager may hedge. The useful lessons are about sizing, which a household controls.
- Measure the loss in money you own. Multiply your largest position's weight by a plausible loss and your portfolio value. At 25% of a €20,000 portfolio, a 30% fall takes €1,500 off the account before other holdings move.
- Check the driver, not the ticker. A storage maker and a chip foundry can depend on the same spending cycle, as Seagate and TSMC did here.
- Test the hedge in the fall you fear. The gold miners were up while storage hit its low, then ended 11.9% down; a holding that diversifies day to day can still lose in the same seven weeks.
- Check what would force you to sell. Compare the stressed balance and your cash reserves with spending and debt payments you cannot postpone. A large holding can interrupt compounding if you have to sell it at a loss.
Our review of Tepper's Appaloosa shows a different Q3 2026 book that added chips, and Third Point's Q3 filing one with a published return record. Our study of falling together shows how correlations change in bad months.
FAQ
What did Stanley Druckenmiller buy and sell in the third quarter of 2026?
Duquesne Family Office's 13F for 30 September 2026, filed on 16 November, shows TSMC cut from 589,680 to 32,700 shares and Seagate raised from 122,000 to 341,790 shares, plus calls on 98,000 more. New positions include Newmont, Barrick and Agnico Eagle, Western Digital, Palantir and SpaceX. Alphabet, AMD, Lam Research, Teva, both airlines and all calls on the S&P 500, Russell 2000 and Brazil ETFs are gone.
What is Druckenmiller's largest position?
Natera, the genetic-testing company, accounted for 29.2% of the filing table's total on 30 September 2026, up from 16.6% in June. The filing's dollar scale is uncertain, so this is a relative weight, not a share of Duquesne's total assets. Almost all of that rise came from the share price, which went from $271.45 to $413.31; Duquesne held 0.8% fewer shares. By 18 November the price had fallen back to $289.50.
Did Druckenmiller sell out of AI in 2026?
Not from the disclosed shares. He cut TSMC and sold AMD, Lam Research and Alphabet, but raised Seagate and added Western Digital and Palantir. Seagate's daily returns correlated 0.52 with TSMC's over the year to September and 0.69 in the seven weeks after. The two names still had meaningful exposure to the same hardware spending cycle.
What return has Duquesne Family Office made?
It publishes none. The last reported figure is for Duquesne Capital, the fund he closed in 2010: Bloomberg reported that its returns had averaged 30 percent a year since 1986, with no annual series. So no geometric return, Sharpe or Sortino ratio can be compared with MSCI World.
Can a household copy Druckenmiller's 13F?
Not usefully. The filing arrived 47 days after the date it describes; by then the September shares, held unchanged, were 15.4% lower. It omits the shorts, futures, currencies, bonds and non-US shares that Druckenmiller has said he uses, including a short bond position.