Research grounding

Walk-forward logistic model (ml_logit)

The app's learned component: logistic regression on returns/RSI/ATR/EMA features, refit at each day boundary on prior days only, labels gated by a cost-aware deadband. Starts UNVALIDATED like everything else. Deliberately simple per arXiv:2603.01820's finding that simple models hold up better on financial data.

News sentiment lexicon

Rule-based word matching over RSS headlines. This is NOT ML and is not presented as the 'AI' part of the app; it is a transparent heuristic veto layer.

Paper: MBOFormer/MBOFusion predict short-horizon trend from level-3 order book events.

Here: Requires market-by-order (L3) data. No free Indian L2/L3 source exists without a brokerage API. Documented as a future adapter; nothing in this app claims LOB-derived signals.

Paper: Regime-robust parameter selection; gradient-boosted trees beat deep tabular models; rank-aggregated ensembles.

Here: We classify intraday regime (volatility percentile + index drift) and gate signals per regime. We do NOT run their Optuna cross-regime search yet; the regime gate is a first-order adaptation.

Paper: Volatility-targeted Sharpe as the evaluation objective; simple models often beat fancy ones; EWMA ex-ante vol.

Here: Position sizing uses EWMA vol-targeting; the validator reports risk-adjusted stats rather than raw accuracy.

Paper: PCA common factors + network VAR on idiosyncratic components for co-movement.

Here: Our 'Idiosyncratic Momentum vs NIFTY' signal uses a single market factor and rolling beta residuals. This is an adaptation hypothesis, not FNIRVAR.

Paper: Falsification of 14 OHLCV intraday momentum families after realistic costs; five validation gates incl. a friction floor.

Here: Our validator implements its gates (t-stat, min trades, net-of-cost positivity, fold consistency, permutation p) plus the friction-floor check. ORB is included precisely to be re-tested, and stays UNVALIDATED until it passes on NSE data.

Paper: Deep RL for intraday decisions.

Here: Offline RL needs a training pipeline and far more data; not wired in v1.

Paper: Expanding-window walk-forward, information-set discipline, interpretable signals, realistic cost model.

Here: Signals carry human-readable reasons; validation is expanding-window with causal features only; costs are modeled, not ignored.