Evidence
The main results, each with the type of evidence behind it. Values are updated by hand and carry date, method and status.
How to read this page
| Evidence type | Meaning |
|---|---|
| Backtest | Historical simulation of rules |
| Out-of-sample | Data not used in selection |
| Forward | Run after the rules were frozen |
| Simulation | Live market, simulated account |
| Real account | Actual broker results |
| Technical | System behaviour, not profit |
Each result states period, instruments, cost model, trade count, drawdown, method and validation status where they apply. We publish no projections.
Three terms appear throughout. R is a unit of risk, so +1 R is a gain equal to the initial risk of a trade. A book is one strategy or variant with its own rules and trade record. Return to drawdown divides total return by the largest peak-to-trough loss.
Rule-based portfolio (historical test)
| Field | Value |
|---|---|
| Period | January 2020 to 1 October 2026 |
| Instruments | NQ and ES |
| Composition | 53 rule-based books, equal weight, one unit per book |
| Trades | 70,881 |
| Costs | Commission and slippage included, about $14 round turn on NQ |
| Sharpe | about 2.15 (daily, annualised) |
| Return to drawdown | about 23 |
| Maximum drawdown | 277 R |
| Positive years | 7 of 7 (2026 to 1 October) |
| Positive months | 57 of 82 |
| Evidence type | Historical test of selected strategies |
Read this before using these figures.
- Most individual strategies were tested with separate in-sample and out-of-sample periods. The portfolio aggregates the historical results of the selected strategies, so the choice of the set was made with hindsight.
- About three quarters of the return comes from the long side in a rising market, and no detrended benchmark exists for the book.
- This is an NQ result. The 9 ES books contribute about 2% of the total.
- 2025 was the weakest full year, at about half the level of 2021–2024.
- R figures use one unit per book across overlapping books, so they are not account results.
- Excluded are neural books, native strategies and one family whose fill order cannot be resolved from minute data. Including that family gives a higher figure, which we do not use.
- Roll weeks remain flagged, so a slight overstatement is possible.
Historical test
53 rule-based books, cumulative result and drawdown (R)
Portfolio diversification (historical test)
| Field | Value |
|---|---|
| Effective exposures | about 12 (55 books, October 2026) |
| Macro bets | 12 (36 strategies, June 2026, hierarchical clustering) |
| Mean daily correlation | about 0.13 |
| Mean correlation on the worst 10% of days | close to zero |
| Limit | Days without a trade count as zero, which lowers measured correlation. Effective exposures measure diversification and are not a count of proven edges. |
Point-in-time review (historical test)
| Field | Value |
|---|---|
| Date | 6 October 2026 |
| Scope | 58 rule-based books |
| Unchanged | 44 |
| Change in total result | −8.4% |
| Limit | Some strategies lost up to about 80% of their expected profit while others gained, so the total hides large differences between books, yet the overall result was only 8.4% lower |
Strategies on simulation accounts (simulation)
| Field | Value |
|---|---|
| Period | 10 July to 3 November 2026 |
| Accounts | 27 simulation accounts |
| Trades | 7,380 |
| Win rate | 42.2% |
| Profit factor | 1.30 |
| Payoff | 1.76 |
| Average per trade | +$44 |
| Result | +$323,003 |
| Maximum drawdown | −$105,805 |
| By month | July +$12,438, August +$13,961, September +$42,203, October +$175,033, November to 3 Nov +$79,368 |
| Entries and exits | Entries are trades that were actually executed, and exits follow the strategy rules. For the 10.5% of trades that ended outside the strategy rules, the rule-based exit was computed from second-level data. |
| Position size | Set by the system and differs between accounts |
| Evidence type | Simulation |
| Limit | Not real capital. Simulation fills orders more favourably than the market, and computed exits assume a fill exactly at the stop or target level. October delivered most of the result. |
Market research (descriptive)
| Finding | Value | Method |
|---|---|---|
| Session size forecast | Brier skill +0.14 NQ, +0.17 ES, positive 7 of 7 years (2026 partial), AUC about 0.7, where 0.5 is chance | Walk-forward, 2020–2026, about 1,690 days per market |
| Direction forecast | AUC 0.52 to 0.54, interval includes no skill | Same study |
| Volatility transfer | 50–56% against 33% by chance | 2020–2026, about 1,650 days, stable in both halves |
| Volume pace | Remaining range about 0.7× against 1.3× of norm, slowest against fastest fifth on NQ, correct sign 28 of 28 | 2020 to May 2026, measured before the clean corpus |
| Reproduced session frequency | 80.7% on ES, 95% CI 79.0 to 82.3, n 2,148 | 2010–2026. A random path gives a similar number, so this is a description and not a signal. |
Drawdown forensics, Q3 2026
| Field | Value |
|---|---|
| Period | 4 August to 14 September 2026 |
| Measure | −246.5 signal R across the live book, −89.9 R as independent bets |
| Attribution | Expectation +59.6, market path −214.7, residual −43.6, classic execution +22.0, trades outside the backtest +16.2, new legs and copies −86 R |
| Reading | The classic book's execution was not the cause. The book-level interval of −291 to +177 R is inconclusive, so edge decay can be neither proven nor excluded. |
| Status | Experimental |
Forensics
Drawdown attribution, 4 August to 14 September 2026 (signal R)
Technical results
| Measure | Value | Date |
|---|---|---|
| Trades reproduced with zero value mismatches on pinned data | 87,260 | 9 Sept 2026 |
| Mixed-contract bars in the live minute archive | 0 of about 11.0 million | 19 Oct 2026 |
| Median internal cycle outside heavy work windows | 16.5 → 2.77 s | 20 Oct 2026 |
| Strategy providers with native fidelity at migration | 52 of 52 | July 2026 |
| Historical trades mirrored with parity checks at cutover | 254,425 | 10 July 2026 |
| Direction splits matching parent, trade by trade | 43 of 43 parents | 3 Sept 2026 |
| Fields of an AI brief recomputed from raw data | 253 | 4 Oct 2026 |
On a rewritten live archive the median match in the reproduction test was 92.4%, which shows that identity holds on pinned inputs.
Withdrawn results
| Result | Reason |
|---|---|
| Early setup results before the May 2026 fix | Look-ahead in day-type classification |
| Earlier projections | No longer valid |
Real account results
Real account results are not published at this time. If that changes, they will be marked as broker-verified and shown separately.
Important notice
Backtested and simulated results are hypothetical and have inherent limits. Past results do not guarantee future results. Trading futures involves substantial risk of loss.