Predictors

Market intelligence for Indian equities — from raw NSE ticks to point-in-time predictors, ML return forecasts, and Lean backtests.

9.2M daily candles, 25 years

5,224 NSE companies tracked

21 predictor families in the registry

1:1 library to REST mapping

One pipeline, four layers

Kite Connect crawlers pull every NSE instrument and 25 years of daily candles into Postgres. Financial statements are processed into point-in-time metrics — a value for any month uses only data that was knowable then, with a configurable reporting lag. Predictor panels feed an ML pipeline that ranks stocks monthly; QuantConnect Lean turns those ranks into auditable backtests.

  crawlers            postgres              compute               serve
┌───────────┐   ┌──────────────┐   ┌───────────────┐   ┌──────────────┐
│ instruments│   │ instruments  │   │ metric        │   │ FastAPI      │
│ historical │──▶│ daily_candles│──▶│ registry       │──▶│ 1:1 REST     │
│ (Kite API) │   │ company_info │   │ (@metric,     │   │ + Python     │
└───────────┘   │ financial_   │   │  dual-tag)     │   │   library    │
                │   data      │   │ panels/*.csv   │   └──────┬───────┘
                └──────────────┘   └───────┬───────┘          │
                                          ▼                  ▼
                                   ML pipeline          dashboards /
                                   (rank stocks)       notebooks / Lean

The predictor registry

Every predictor is a registered function carrying two tags: a standard financial category (for humans) and an ML feature-group alias (for models). Adding one predictor is one decorator — the panels, the API, and the registry pick it up automatically.

ML group	Category	Predictors
momentum	momentum	mom1m · mom6m · mom12m · mom36m · chmom · indmom
valuation	value	bm · bm_ia · ep · sp · dp · dy · cashpr · rd_mve
profitability	profitability	operprof · ps
growth	investment	agr · invest · chinv · nincr
leverage	investment	lgr · chcsho
size	size	mvel1

One API, two access paths

All logic lives in a Python library. The REST layer is a thin wrapper — every endpoint maps to exactly one library function, with zero logic of its own. Point-in-time guarantees travel with the call: asof_lag controls how stale financial statements are allowed to be.

Endpoint	Library function
GET /tickers	list_tickers(session, sector?, industry?)
GET /tickers/{code}/daily	get_daily_candles(session, code, start?, end?)
GET /tickers/{code}/metrics	get_metrics(session, code, metrics?, asof_lag?)
GET /panel/{metric}	get_panel(metric, start?, end?)
GET /metrics	list_metrics()

Backtest results

Monthly stock ranking on 447 engineered features against the Fama-French 6-factor benchmark, evaluated in QuantConnect Lean.

Model	Approach	Cum. return	Sharpe	Info. ratio
OLS-all L2	Ridge on 447 features	1.362	0.792	0.812
OLS-all Huber+EN	ElasticNet-Huber, sparse	1.154	0.724	0.130
GBRF Huber	Gradient-boosted, robust loss	1.210	0.757	0.303
OLS-3 Huber	Fama-French baseline	1.084	0.689	-0.105

Best alpha vs FF6: +0.26%/month (OLS-all L2). Predictive R² reported per model in the full research report.

Built with

Python 3.13 · uv workspace · Postgres 16 · pandas · SQLAlchemy · FastAPI · scikit-learn · QuantConnect Lean · pytest

Talk to us about it