Before the narrative, the scorecard. Each figure below is read directly from the run artefacts; the derivations are explained in the sections that follow. The first column is what a return-seeking allocator looks at. The second is what a risk officer looks at. Both columns belong to the same book.

Indexed to 100 at the open of 2 January 2024, the 10-name book finished 1 October 2026 at 616.6. The benchmark finished at 161.6. Peak wealth on the strategy reached 728.1 on 22 June 2026, before the summer drawdown, which is why the ending point sits 15.3% below the high-water mark and why we report Calmar alongside CAGR. In dollars: $200,000 compounded to $1,456,219 at the peak and $1,233,163 at the close of the window.

The curve is plotted on a log scale so that the slope reads as a growth rate. Three things stand out. The first year was the steepest: 2024 compounded 161%. The 2025 spring correction cost 37% peak-to-trough and was fully recovered by July. The spring 2026 rally (April +37.0%, May +24.9%) took the book to its all-time high; July 2026 (−25.7%) gave about half of that back. The strategy has never closed a session behind the benchmark on this growth-of-100 basis, and has not been below its starting value since its second session.

On every horizon longer than a quarter, the strategy is ahead: one-year +35.2% versus +14.3%, year-to-date +59.5% versus +12.0%, six-month +56.6% versus +16.6%. Daily Sharpe (1.62 vs 1.21), Sortino (2.63 vs 2.01) and Calmar (2.44 vs 1.01) all favour the book. On a monthly basis the Sharpe advantage narrows to 1.51 against 1.46, which tells you the edge is a daily-compounding edge in a high-volatility series rather than a smooth monthly one.

The honest caveats sit in the same table. Maximum drawdown is −38.6% against −19.0% for the benchmark. Daily volatility is 48.5% against 15.6%. The latest three months are −5.0% against +2.4%—momentum lagged while the index ground higher after the July rotation. A strategy that concentrates in the ten strongest names will give back more when leadership rotates, and it did in the summer of 2026, for the second summer in a row.

The strategy beat the benchmark in every calendar year of the sample. 2024 was the standout: +161.1% against +24.0%, carried by an AI-infrastructure leadership regime that the 252-day leg of the score held onto for most of the year. 2025 was the hard year: a −36.9% drawdown from mid-February to early April, a recovery by July, a strong September (+19.7%), then a November payback (−17.0%). The book still finished +48.0% against +16.4%. 2026 through 1 October is +59.5% against +12.0%, carried by April and May and partly surrendered in July.

The pattern that matters for sizing is the width of the annual outcomes: 161%, 48% and 60%. The worst full year was still twice the benchmark’s best. A CAGR of 94% is an average of very different years, not a run-rate.

A regression of the book’s daily returns on the benchmark gives a beta of 1.98 and a daily correlation of 0.63. The high beta is structural: ten equal-weight positions in the fastest-rising stocks in the market will always carry more market exposure than the index. What the regression does not explain is a residual of about 41% per year, annualised from daily data. Even after charging the book for two units of market risk, roughly two-fifths of its annual return is not accounted for by the index.

Capture ratios tell the same story from a different angle. On days the benchmark rose, the strategy rose 2.50 times as much on average. On days the benchmark fell, it fell 2.07 times as much. The up-capture exceeds the down-capture by 0.43; that gap, compounded over 690 sessions, is the excess return. Note that the down-capture is still greater than two. This is not a hedge, and it has never behaved like one.

Average drawdown on the 10-name book lasted 20.8 trading days and ran about 6.7%. The worst peak-to-trough was 38.6%, from the 22 June 2026 high to 29 July 2026—37 calendar days to lose more than a third of the book. The benchmark’s worst was 19.0%, in the April 2025 sell-off. The strategy has now been below its June high for 101 days and sits 15.3% under it as this edition goes out.

The three large episodes have different shapes. Spring 2025 was fast down (49 days) and moderately fast up (101 days to a new high). Winter 2025–26 was a slow grind: 152 days from the October peak to the March trough, then a 32-day recovery once the April rally began. Summer 2026 was the fastest decline of the three and is not yet recovered. An allocator should expect, from this sample, roughly one drawdown of 30% or more per year and a recovery time measured in months rather than weeks.

What caused the summer 2026 drawdown

The run decomposes every day of drawdown into a benchmark component (what the market did), a selection component (what the chosen names did relative to the market) and a sizing component (what equal weighting did relative to the market’s own weights). From 22 June to 29 July 2026, the benchmark component was almost nil—the index was flat in July. Selection explains roughly 90% of the decline and sizing the remaining 8%. The names that did the damage were Bloom Energy, Western Digital, SanDisk, Applied Optoelectronics and TeraWulf: two of them were the book’s largest winners of the year, which is what a momentum reversal looks like from the inside. The leaders fell together because they had risen together.

Across the full sample the same decomposition holds: selection accounts for about 88% of all drawdown-days, the benchmark for 6% and sizing for 6%. The path risk in this strategy is a stock-selection risk, not a market-timing risk. It cannot be hedged away with index futures without also hedging away the return.

Across 438 rolling one-year windows, the strategy’s trailing twelve-month return was positive in every one of them. The lowest reading was +19.6%. The benchmark’s lowest was −4.2%. The strategy beat the benchmark in 63% of those windows. Over rolling three-month windows the picture is noisier, as it should be: positive 72% of the time and ahead of the benchmark 60% of the time, with the latest print +2.1% against +2.6%.

The rolling six-month Sharpe ratio has ranged from −0.58 (spring 2025) to 4.68 (autumn 2024). It is currently 1.75, below the benchmark’s 2.53 over the same six months because the index had a quiet, steady summer while the book had a violent one. The full-sample daily Sharpe of 1.62 is an average across regimes, not a constant, and the strategy’s rolling Sharpe exceeded the benchmark’s in only 51% of windows. The compounding edge comes from the magnitude of the good windows, not from winning most windows.

2024 compounded +161.1% with only three losing months. 2025 finished +48.0% despite February, March, November and December all negative. 2026 year-to-date through 1 October is +59.5%, carried by April (+37.0%) and May (+24.9%), then given back in July (−25.7%). Up months average +13.1%; down months average −10.4%. Twenty-four of thirty-four months were positive and twenty-four of thirty-four beat the benchmark. That asymmetry, more up months and bigger up months, is the premium.

Two seasonal observations, offered as observations rather than rules. Both large year-end drawdowns (December 2024, November–December 2025) came after a strong autumn, consistent with momentum crowding into year-end and unwinding. And the single worst month, July 2026, arrived in a month the index finished flat. The benchmark heatmap underneath shows how little of either pattern was market-driven.

A portfolio return can be luck. A signal that ranks a whole universe correctly, month after month, is harder to dismiss. At each rebalance the run records the Spearman rank correlation between the momentum score and the subsequent return of every name in the 75-stock screened universe (the information coefficient, or IC), together with the return spread between the top and bottom fifths of the ranking. Both are measured at four forward horizons: 5, 10, 21 and 63 trading days.

At the 21-day horizon—the one the book is actually rebalanced on, and the one pre-registered as the primary endpoint before the run—the mean IC over 33 months is 0.087 with an IC information ratio of 0.43. The top quintile returned 6.6% per month on average against 2.8% for the bottom quintile: a 3.7 percentage-point monthly spread. A moving-block bootstrap of 1,000 resamples places the 95% confidence interval for the IC at 0.017 to 0.156, with 99.5% of resamples positive and a p-value of 0.006. The quintile spread’s interval is 0.5 to 7.3 percentage points, p-value 0.010. Both clear the pre-registered threshold. For an equity factor, an IC near 0.09 at monthly frequency is a strong reading; single-factor ICs commonly cited in the practitioner literature sit between 0.02 and 0.06.

The shorter horizons are weaker and, after a Benjamini–Hochberg adjustment for testing two metrics, the 5- and 10-day quintile spreads do not reach significance, though their ICs do. The 63-day horizon has the highest hit rate (68% of windows positive) but only ten independent evaluation periods, so the run correctly labels it “limited history” and declines to attach a confidence interval. We report the exploratory horizons for completeness; the claim rests on the 21-day endpoint.

The 2025 anomaly

Split by calendar year, the 21-day IC was 0.148 in 2024, 0.003 in 2025 and 0.119 in 2026. In 2025 the score had essentially no ability to rank the whole universe—yet the portfolio still returned +48%. The reconciliation is that the strategy does not trade the whole ranking; it trades the top ten. A factor can lose its monotonic relationship across the middle of the distribution and still identify the extreme winners. The temporal-robustness tables in the run show the same pattern: the 2025 calendar sub-period spreads are flat at the 5-, 10- and 21-day horizons and only mildly positive at 63 days, while the 2024 and 2026 sub-periods are strongly positive throughout. We flag this because it is the kind of detail that gets lost in a headline IC, and because a year in which the signal goes quiet is exactly when a momentum book is most exposed to a reversal.

The blotter contains 464 fills across 34 rebalances: 235 buys and 229 sells, an average of 13.8 tickets per rebalance. Lot-matched, that is 309 round trips with 211 wins and 98 losses: a 68.3% hit rate. Measured at the position level (229 exits, including rebalance trims), the average winning exit returned +44.7% and the average losing exit −10.6%, a payoff ratio of 4.2. Gross realised gains were $1.20 million against gross realised losses of $0.47 million, a profit factor of 2.6. Mean return per exit was +29.5%; median holding period 57 calendar days, mean 78.

The distribution is what you would expect from a trend strategy: 48 exits returned more than 50%, 27 more than 100%, and only three lost more than 30%. The ten largest realised gains account for 56% of net realised P&L. That concentration is the mechanism, not a flaw. Equal weighting and monthly re-ranking mean a losing position is trimmed or removed within a month while a winning one is kept until its rank fails. The book is designed to let the right tail run and cut the left tail short, and the blotter confirms that it did.

Turnover and costs

On average 4.2 of the 10 names change at each monthly rebalance (minimum 1, maximum 8; no rebalance left the book unchanged). One-way turnover averages 44% of net asset value per month, or about 530% annualised. One hundred and six distinct names passed through the book in 34 months. Modelled costs totalled $5,659 on $18.5 million of notional traded—3.1 basis points of notional, or roughly 0.30% of average NAV per year. That is a light cost model. A live book of this size would pay more in spread on the smaller names, and a materially larger book would move prices in some of them; we address that in the closing note.

Summing each name’s daily weight times its daily return, scaled by the book’s wealth at the time, gives an additive contribution to the growth of 100. Of the 103 names that contributed, 68% added value and 32% subtracted. The ten largest contributors added 404 points of the 529 additive points generated—about three-quarters—and four of them are memory and storage: SanDisk (+65), Micron (+59), Western Digital (+49) and, from a different corner, AppLovin (+55). Strategy (MSTR), IREN, Bloom Energy, Palantir, Intel and Carvana fill out the top ten.

The bottom of the table is instructive. Planet Labs (−25), Applied Optoelectronics (−25), NuScale (−23), TeraWulf (−18) and Nebius (−13) were all 2026 entries that reversed within one or two months of being bought. Applied Optoelectronics appears on both the loser list and the current holdings list: it was sold at the August rebalance and re-entered in September when its rank recovered. That is the system doing what it is supposed to do, without memory and without ego.

The October rebalance still lives in semiconductors and storage. SanDisk (SNDK) leads the ranking by a wide margin, with a momentum score more than twice that of the next name; it has been in the book for eight consecutive months. Micron (MU, nine months) and Western Digital (WDC, ten months, the longest current tenure) sit behind it. ASE Technology (ASX), Dell (DELL), Roivant Sciences (ROIV) and Applied Optoelectronics (AAOI) carry over from recent months. Three names are new this month: ImmunityBio (IBRX), a biotech that has climbed into fourth place; Advanced Micro Devices (AMD); and United Microelectronics (UMC), which makes four semiconductor manufacturers plus two storage names in a ten-name book. Each position is 10% of the book. Sixty-eight names cleared the screen; only these ten are held.

The book is concentrated by design. Eight of the ten holdings are in the semiconductor, storage or hardware supply chain and will respond to the same AI-capex cycle. Two biotechs provide the only diversification. A reader who wants to know what this strategy owns can summarise it in a sentence: the memory cycle, the foundries, and whatever else is going up fastest.

Indexed to 100 at the open of 2 January 2024, the 12-name book finished 8 September 2026 at 443.3. The benchmark finished at 162.1. Peak wealth on the strategy reached 573.8 before the 2026 mid-year drawdown, which is why the ending point sits below the high-water mark and why we report Calmar alongside CAGR.

On every horizon longer than a quarter, the strategy is ahead: one-year +44.2% versus +17.8%, year-to-date +37.5% versus +12.1%, six-month +38.9% versus +12.9%. Daily Sharpe (1.43 vs 1.23), Sortino (2.31 vs 2.05) and Calmar (1.88 vs 1.04) all favour the book.

The honest caveats sit in the same table. Maximum drawdown is −39.4% against −19.0% for the benchmark. Daily volatility is 46.9% against 15.7%. The latest three months are −6.7% against +3.9%—momentum lagged while the index recovered. A strategy that concentrates in the strongest names will give back more when leadership rotates, and it did in the summer of 2026.

The 12-name book is the production sleeve because it balances concentration and diversification. The same factor, screen, and window was also run at 10 and 20 longs. Every long-only sleeve beat the benchmark on total return, CAGR, Sharpe and Calmar. A tighter 10-name book compounded faster (CAGR 94.8%, total return 498%). A 20-name book was slower (CAGR 72.3%) but still more than tripled the index. The premium does not depend on an exact position count.

A mixed 10-long / 2-short experiment in the same series did not. It finished barely above water, with a 75% drawdown. Production stays long-only for that reason.

Average drawdown on the 12-name book lasted 17 trading days and ran about 6.3%. The worst peak-to-trough was 39.4%. The benchmark’s worst was 19.0%. That is the price of concentrating in high-momentum names, many of them semiconductors and hardware in this sample.

2024 compounded +105.2% after the January start. 2025 finished +33.1% despite sharp February and November paybacks. 2026 year-to-date through 8 September is +37.5%, carried by April +33.3% and May +25.3%, then given back in July (−19.5%). Up months average +12.9%; down months average −6.8%. That asymmetry is the premium.

The latest rebalance still lives in semiconductors, storage, and a handful of non-tech names that cleared the same score. SanDisk (SNDK) leads the factor. Micron, Western Digital, Intel, Dell, Applied Optoelectronics, ASE and Coherent sit behind it. Warner Bros. Discovery, Roivant, Bloom Energy and NUAI complete the twelve. Each is 8.33% of the book. Sixty-six names survived the screen; only the top 12 are held.

On the same 2.7-year path, the 12-name QS book turned 100 into 443. The 20-name XGB book turned 100 into 286. The benchmark finished at 162. Adding a gradient-boosted model and three fundamental scores did not beat a clean momentum rank in this window. It did change the shape of the ride: XGB’s latest three months are +3.7%, in line with the index, while momentum gave back 6.7% as leadership rotated. Maximum drawdown is similar (−36.0% vs −39.4%). Sharpe is 1.27 vs 1.43 vs 1.23 for the index—XGB still clears the benchmark on a risk-adjusted basis, just not by as much as the factor book.

Indexed to 100, the 20-name XGB book finished at 286.2. That is +186% versus +62% for the benchmark—124 percentage points of excess return. Calmar 1.33 still beats the index 1.04. The model earns its keep against a passive book. It does not earn a claim of superiority over the simpler momentum rank on this sample.

Five finished XGB specifications were compared, all using Rule of 40, FCF and Lynch screens, all long-only, all the same dates. The 20-name book on the top 100 names by volume is the best of them (Sharpe 1.27, total return 186%). Cutting to 10 names on that same tight universe still returned 167%. Widening the screen to the top 1,000 names hurt: 20 names made 107%, 10 names made 66% (barely above the index) and a 30-name book on that wide universe made 77%. Liquidity concentration matters more for this model than adding names.

2024 compounded +44.3%. 2025 was the best calendar year at +48.3%, carried by a spring-to-September run (April through September all positive, September +19.2%) after a hard February (−12.1%). 2026 year-to-date through 8 September is +28.8%, with April +20.6% and May +19.8% doing the work after a weak first quarter. Unlike QS momentum, July 2026 is only −2.0% here, not −19.5%. That is the quality overlay showing up in the tape.

Joby Aviation leads the predicted 21-day return. Micron is the overlap with the QS momentum book above. Then Oracle, AMD, a cluster of silver and gold miners (Coeur, Hecla, Barrick, First Majestic), Robinhood, Marvell, NVIDIA, Broadcom, Amphenol, Microsoft, CleanSpark, Hims, Carvana, Alphabet and Palantir. Five percent each. The momentum book is a semiconductor-and-storage concentration. This one is a barbell: mega-cap quality on one side, high-beta miners and unprofitable growth on the other. That is what you get when a tree model is allowed to mix Rule of 40 with raw momentum.

A long stock position needs the asset to appreciate before it books a capital gain. A bull put spread does not. QQQ can rise, move sideways, or even decline moderately, and the position can still keep the full premium, provided QQQ stays above the short put strike at expiration.

That is a useful gap between the directional forecast and what the trade actually needs. We do not have to say how high QQQ will go. The better question is how likely QQQ is to remain above a particular downside level over the life of the trade.

The original analysis compared SPY and QQQ. The system produced the same covered-call recommendation for both. QQQ showed the more attractive relative implied-volatility characteristics in that work. That matters because we are selling option premium. Higher implied volatility usually means more premium in the options. It is not a hunt for the jumpiest ETF. More volatility is also more risk.

QQQ can move larger and faster than SPY because it is more concentrated in technology and growth names. The extra premium is compensation for that extra market risk, not free money. The desk is looking for a balance between premium received and risk assumed.

A defined-width bull put spread can use far less buying power than 6,000 QQQ shares. Capital can stay available for other strategies, reserves or diversification. Capital efficiency is not safety. A defined-risk spread can still lose the entire capital allocated to that position. The advantage is that the loss boundary is known.

The intended entry is 60 spreads at a $0.19 net credit limit, not whatever the market prints. Bid-ask spreads matter when the same structure is traded repeatedly. A few cents look small until they are scaled.

All 60 spreads are exposed to essentially the same QQQ event. If QQQ gaps lower, they can move against the book together. Size is not justified by a high estimated probability on each contract. The position is one portfolio exposure: maximum dollar loss, portfolio percentage at risk, delta, volatility, and the chance of a sudden gap.

The interesting part of this trade is not the 23% headline. It is that the algorithm recommended one structure and the desk selected another. That is not a system failure. It is the infrastructure working as decision support rather than a black box.

The system identified a bullish trend, low implied volatility and an income objective, and suggested a covered call. The desk then added capital efficiency, defined downside, QQQ option pricing, directional conviction, desired exposure and portfolio construction. The resulting order is 60 QQQ bull put spreads at a $0.19 limit credit.

We do not treat the future of this stack as asking a model what to trade today. Algorithms generate structured evidence. Risk systems set the boundaries. Execution systems implement the trade. Human oversight still asks whether the complete decision makes sense.