SHAP for Trading Models: Opening the Black Box

OBSERVED: A NIFTY XGBoost model signals BUY. The desk asks "why?" If the answer is "the model said so," nobody sizes the trade. SHAP (SHapley Additive exPlanations) turns that black box into a per-feature reason: "EMA spread +0.4, GEX +0.3, RVOL −0.1 → net BUY." Now the desk can trust or override.

SOURCE: Lundberg & Lee (2017) SHAP, a game-theoretic feature-attribution method. Applied to gradient-boosted trees via TreeSHAP. Used in the NIFTY XGBoost corpus (feature-importance + SHAP studies) and the BTC LightGBM shadow trader.

DERIVED: How to read SHAP on a trading signal and why attribution decays.

1. What SHAP Actually Gives You

For one prediction, SHAP assigns each feature a value: how much it pushed the output up (+) or down (−) from the baseline. Sum of SHAP = model output − baseline.

On a BUY signal:

EMA50-200 spread : +0.42
Gamma exposure   : +0.31
RVOL             : -0.08
IV rank          : +0.05
=> sum pushes prob from 0.50 to 0.70 (BUY)

You can read the trade like a sentence.

2. Why Traders Need It (Not Just Researchers)

Trust: a reason beats "model said so."
Override: if SHAP shows the signal is driven by a stale feature, skip it.
Compliance: SEBI/risk desks want explainability; SHAP is the standard.
Debug: a feature with huge SHAP but no economic meaning = leak or bug.

3. SHAP vs Feature Importance

Method	What it tells	Problem
Gain/imp	global, averaged	hides per-trade direction
Permutation	global, honest	slow, less local
SHAP	per-prediction	gold standard for explainability

SOURCE: The corpus tracks both global importance (37 files) and SHAP (18 files). SHAP is the per-trade lens; importance is the overall ranking.

4. Reading SHAP on NIFTY

A real signal from the corpus shape:

Trend features (EMA spread, structure) → steady positive on uptrend signals
Flow features (option-chain precursor, RVOL) → spike on event bars
Vol features (IV rank, GEX) → flip sign near expiry

When SHAP flips sign on a normally-positive feature, that is a regime change — worth a hold.

5. Code Sketch

import shap, xgboost as xgb
model = xgb.Booster(); model.load_model("nifty_15m.json")
expl = shap.TreeExplainer(model)
sv = expl.shap_values(X_row)   # one prediction
# sv[0] = per-feature contribution
for name, val in zip(feature_names, sv[0]):
    print(f"{name}: {val:+.3f}")

TreeSHAP is exact for trees — no approximation.

6. SHAP Decay (Feature Drift)

A feature important in 2022 may vanish by 2025. SHAP distribution shifts → the model is trading stale logic. Fix: recompute SHAP on rolling windows; retire features whose contribution collapses.

OBSERVED in corpus: RVOL contribution dropped post-2023 regime; the rolling-ablation study caught it and the model was retrained without it.

7. Common SHAP Mistakes

Reading global mean only — you need per-trade for trust.
Ignoring sign — magnitude without direction is meaningless.
SHAP on leaked features — explains a lie; fix leakage first (Label article).
No baseline — SHAP is relative to the dataset mean; know your baseline.

8. FAQ

Q: SHAP slow?
A: TreeSHAP is fast for XGBoost/LightGBM. On 1 prediction, milliseconds.

Q: Works on LSTM?
A: Harder (no TreeSHAP). Use Integrated Gradients or attention weights instead.

Q: Required by SEBI?
A: Not literally, but explainability is expected for risk desks. SHAP is the standard.

Q: Advice?
A: No. Educational. NISM-Series-XII educator, not SEBI RA.

8. Worked Example: One BUY Signal, SHAP Read-Out

A real NIFTY 15m BUY (prob 0.70), baseline 0.50:

Feature                 SHAP    Direction
EMA50-200 spread        +0.38   uptrend confirm
Gamma exposure          +0.22   dealer support
RVOL (volume)           +0.10   participation
IV rank                 -0.04   vol not extreme
Option-chain precursor  +0.04   mild bid
Sum                     +0.70

Reading: trend + flow drive it; IV is neutral. If next bar RVOL flips negative AND SHAP shows −0.15, that is a warning the participation died — hold, don't add.

DERIVED: SHAP turns "BUY" into "BUY because trend+flow, watch RVOL." The desk can act on that.

9. SHAP Decay Table (Feature Drift)

Feature	2022 contrib	2025 contrib	Action
EMA spread	+0.35	+0.33	keep
RVOL	+0.18	+0.06	decayed → retrain
IV rank	+0.05	+0.04	keep
Old news feat	+0.12	+0.01	drop

The corpus's rolling-ablation caught RVOL decay; model retrained without the stale news feature. SHAP distribution shift = your early-warning system.

10. SHAP for Risk Sizing

SHAP magnitude can size risk: a signal driven by one dominant feature (e.g. GEX +0.5 alone) is fragile — size it small. A signal with several moderate features (EMA +0.3, flow +0.2, vol +0.1) is robust — size it full. The corpus's "soft-risk sizing policy" does exactly this: confidence spread across features → position size.

11. Global vs Local SHAP

Local (one prediction): "why THIS trade" — for trust/override (what we did above).
Global (all predictions averaged): "what the model relies on overall" — for audit/retire.

Global SHAP on the corpus showed EMA-spread + GEX = 70% of average contribution; tiny features (<1%) got retired. Local SHAP is for the desk; global is for the researcher. Both from the same shap.TreeExplainer.