SAMPLE REPORT — illustrative deliverable. Figures are real results from QuantAssay internal research, not a client submission.
REPORT  QD-0001
ISSUED  2026-08-01
SPEC HASH  a7f3c9e1

Strategy Autopsy — NQ/ES dislocation short

You asked whether the edge is real. It is. That is not the same as it being safe to trade, and the difference is where this report spends its time.

Verdict
Paper desk: approved. Funded eval: denied.
Realistic expectancy — 1 NQ
$33,129 / yr
Survives $1k trailing drawdown?
No
The signal is directional and it is not noise — it cleared 3 of 3 out-of-sample folds. But 79% of the entire P&L came from 2 of 11 months, and a $6,924 peak drawdown ends a $1,000-trailing evaluation long before the good months arrive. Trade it on your own account. Do not put an evaluation fee behind it.
// 01 — what you submitted

Your idea, in your words

"When NQ runs way ahead of ES — like the spread blows out well beyond normal — I want to short NQ and ride it back to fair value. Feels like it works best when the dislocation is extreme, not just mildly stretched."

Before testing anything we froze that into a spec you approved. Once frozen, the parameters below could not change — every result in this report was produced by exactly this configuration.

ParameterFrozen valueWhy
InstrumentsNQ vs ES (front month)Automatic contract rolling on CME expiry
Signalhedge-ratio spread, W200 rolling window200-bar window to define "normal" spread
Triggerz > 3.5, transition onlyFires on first crossing, not every bar it stays stretched
DirectionSHORT NQYour stated direction; the long side was tested separately
SessionAll sessionsYou did not restrict; we tested splits anyway
Bracketsderived from training data onlyTP/SL from train-fold MFE/MAE percentiles — never from test
Costs$1.00 round-turn + spread crossingShorts enter at bid, exit at ask
// 02 — how it was tested

Eleven months of out-of-sample, walk-forward

Tick-by-tick simulation. The engine sees data sequentially and strategies are only ever handed history — never a bar that has not happened yet. Brackets are re-derived on each training fold and applied blind to the following test fold.

151
OOS trades
62.25%
Win rate
3 / 3
Folds profitable
10.74
Calendar months OOS

On its face that is a strong result, and plenty of vendors would stop here and sell you the number. We then spent the rest of the engagement trying to break it.

// 03 — the risk department

Five attempts to kill it

1 — Concentration: does it rest on a couple of lucky months?
Failed

Yes, badly. March 2026 alone produced $12,907 of the $29,650 total — 43.5%. Add December 2025 and just 2 of 11 months account for 79.0% of every dollar.

Strip those two months and the remaining nine annualise to $8,561, not $33,129. Four of eleven months lost money.

2 — Parameter stability: does it want the same trade each fold?
Failed

No. The take-profit the training data asked for ranged from 15.44 to 63.00 points — a 4.1× spread — and the risk profile inverted completely across folds.

Fold 2 wanted a wide target and tight stop (TP/SL 2.43, 30-minute hold). Fold 4 wanted the exact opposite (TP/SL 0.37, 240-minute hold). A stable edge does not change its mind about what it is. This looks like a signal being re-fit to each regime rather than one persistent effect.

3 — Prop-firm drawdown: would it survive an evaluation?
Failed

No. Peak-to-trough drawdown of $6,924, with a single worst day of −$3,458. Against a $1,000 trailing drawdown you are eliminated many times over, and the worst day alone ends the account on its own.

4 — Direction: is the short side actually the edge?
Survived

Yes — your instinct was right. We ran the mirrored long (z < −3.5) on the same data. It managed only 1–2 of 3 folds positive versus 3 of 3 for the short. The asymmetry is real: NQ dislocating rich mean-reverts more reliably than NQ dislocating cheap.

5 — Resampling: is the result robust to reordering?
Partial

Bootstrapping the monthly P&L (20,000 resamples, seed 20260801) leaves it profitable in 97.3% of draws — the edge is not a coin flip.

But the spread is enormous: 5th percentile $4,421, median $32,237, 95th percentile $64,999. A 15× range between bad luck and good luck means the headline number is close to meaningless as a forecast for any single year.

// 04 — the picture that matters

Where the money actually came from

+13k 0 $12.9k $10.5k JulAugSep OctNovDec JanFebMar AprMay 2025 2026
Monthly out-of-sample P&L, 1 NQ contract, $1.00 round-turn. Two bars carry the strategy. Nine months of trading produced $6,236 between them.
// the honest read

If you had traded this from July 2025 you would have spent five months roughly flat-to-losing before December paid you. Most traders do not survive that stretch psychologically, and no evaluation account survives it mechanically. The edge is real; your ability to still be trading when it shows up is the actual constraint.

// 05 — what we would do next

If this were our capital

  1. Do not run it in an evaluation. At $1k trailing drawdown the maths is settled — a $3,458 losing day cannot be survived. Any eval fee spent on this configuration is a donation.
  2. Investigate what December and March had in common. Both were high-realised-volatility regimes. If the edge is conditional on a volatility state, a regime filter converts a lumpy strategy into a selective one that sits out the dead months.
  3. Fix the bracket instability before adding size. Refit on a rolling basis with a single stable bracket rather than per-fold optimisation, and accept a lower headline number in exchange for one you can actually rely on.
  4. Forward paper for 30 days minimum. The gate is not optional on our platform. Live export unlocks after 30 sessions of forward paper, with daily replay against the frozen spec.
  5. Consider the US-session variant as a comparison. Restricting to US hours gave a 70.2% win rate over 57 trades — fewer trades, cleaner behaviour, but the same drawdown problem remains unsolved.
// 06 — check our work

Provenance

Every number above traces to a file. The monthly series sums to $29,650.25, matching the recorded total exactly; the three fold trade counts sum to 151, matching the recorded OOS count. If you want the raw trade ledger, ask and we will send it.

  • signal  →  hedge|W200|t3.5|all|S  (proper walk-forward, train-only brackets)
  • results  →  proper_walkforward_results.csv
  • method  →  proper_train_select.py
  • OOS window  →  2025-07 … 2026-05  (10.74 calendar months)
  • bootstrap  →  20,000 monthly resamples, numpy default_rng(20260801)
  • fold P&L  →  F2 $5,559.50 · F3 $5,742.25 · F4 $18,348.50
  • fold WR  →  F2 54.17% · F3 47.73% · F4 79.66%