ML Sizing Overlay — validated results¶
NIFTY + SENSEX · rolling-quarter walk-forward · out-of-sample
Our proprietary ML model scores every option leg of a strategy by its loss-risk and turns that into a position size — it down-sizes the risky legs and keeps the clean ones at your base size, trading every leg. Applied as a sizing overlay (not a filter, not a new strategy), across naked and hedged short-premium structures it lifts out-of-sample, risk-adjusted returns materially — and honestly flags the strategies it does not help.
How wide the edge is¶
Across a large synthetic backtest (NIFTY + SENSEX, real 1-minute prices, many strikes / stops / entry times), the sizing beats flat on a rolling-quarter walk-forward in:
| configurations that improve | |
|---|---|
| Intraday short premium (straddle / strangle / iron-fly / condor) | ~93% |
| Positional short premium (overnight / multi-day) | ~76% |
Two worked examples below — one naked-intraday, one hedged-positional.
How De-risk sizes¶
De-risk down-sizes the risky legs and keeps the clean ones at your base — capped at 1×, so it never adds margin and your average exposure is unchanged. The payoff is a lower drawdown and a higher risk-adjusted return on the same capital — loss-reduction, not leverage. The worked examples below compare FLAT vs De-risk on two strategies.
Case 1 — SENSEX short straddle (naked · intraday)¶
base 10 lots · gross · 537 trades
| metric | FLAT | De-risk |
|---|---|---|
| Sharpe (full period) | 1.26 | 1.84 |
| Max drawdown | ₹15.6L | ₹9.3L |
| Return / MaxDD | 1.99 | 3.99 |
| Return on capital (gross) | 81% / yr | 96% / yr |
| quarters beat flat | — | 5/9 ✅ |
De-risk cuts the drawdown ~40% (₹15.6L → ₹9.3L) while lifting Sharpe 1.26 → 1.84 and the recovery factor (Return/MaxDD) 1.99 → 3.99 — same legs, same average exposure, materially smoother equity.
Case 2 — NIFTY iron-fly (hedged · positional)¶
base 10 lots · gross · 384 trades · the protective wings are kept (hedge ratio preserved)
| metric | FLAT | De-risk |
|---|---|---|
| Sharpe (full period) | 1.70 | 2.38 |
| Max drawdown | ₹6.66L | ₹3.71L |
| Return / MaxDD | 3.44 | 6.91 |
| quarters beat flat | — | 6/9 ✅ |
| hedge-ratio drift | — | 0.05 (intact) |
A hedged, multi-day strategy — and the model sizes the whole structure as one unit, so the protective
wings ride along untouched (drift ≈ 0). De-risk lifts Sharpe 1.70 → 2.38 and nearly halves the
drawdown (₹6.66L → ₹3.71L), doubling Return/MaxDD to 6.91.
Why these numbers are trustworthy¶
- Rolling-quarter walk-forward. Not a single train/test split — sizing must beat flat in a majority of independent quarters and not be worse risk-adjusted over the full span. Each leg is scored using only information available at its entry moment (no look-ahead).
- Whole-lot, deployable. Sizing is in whole lots and every leg is kept (floored at 1); it's built for 10+ lots and engages at ≥ 5 base (below that it runs flat).
- Per-strategy calibration. Calibrated once per strategy so the average size equals your base — the uplift is allocation, not leverage.
- Hedge-safe. Hedged structures are scaled at the structure level with a hedge-ratio guard — the model refuses any sizing that would strip your wings.
- Honest by construction. Run on a strategy the signal doesn't fit, the same tool returns "run flat".
Scope — where it does not help (and the tool says so)¶
| strategy | verdict |
|---|---|
| naked short premium — straddle / strangle & their delta-shift / stop variants | ✅ improves (intraday strongest) |
| wide-wing hedges — iron-fly / iron-condor (cheap far wings) | ✅ improves (sized as one unit) |
| tight credit spread (narrow, defined-risk) | ❌ run flat — reward is capped, sizing can't help |
| long / option-buying | ❌ run flat — the model sizes short premium |
Caveats¶
- Gross of costs. Numbers exclude charges; realistic costs are a ~15–16% haircut on gross profit for a high-frequency premium-seller. Net edge is smaller, but the sized version stays ahead.
- Per-strategy calibration required (one-time — the backtest does it for you).
- Built for 10+ lots — sizing engages at ≥ 5 base (below that it runs flat).
- Match the horizon — an intraday strategy must use the intraday model, positional the positional one.
- Validated on NIFTY + SENSEX, 2024–26.
Try it on your own strategy¶
Upload your AlgoTest CSV on the /backtest page in the app, pick the Hold (intraday / positional) —
you get this exact flat-vs-sized (De-risk) comparison, the calibration key to use live (see
Sizing), and a downloadable per-leg tape.