Sentinel — Point-in-Time Market Research OS

Overview Sentinel is a point-in-time market research operating system for Indian equities — not a stock picker, not a buy/sell recommender, not a chatbot. It's research infrastructure designed to make it difficult to fool yourself with a backtest. The Problem Most retail quant projects produce backtests that look incredible and fail completely in live trading, because they unknowingly leak future information — using revised fundamentals, restated macro data, or news that wasn't actually knowable at that point in history. Sentinel treats this leakage risk as the central engineering problem, not an edge case. What Makes It Different - Point-in-time discipline as a sacred rule. Every data access enforces through a single gateway module — never bypassed, never assumed. - Immutable, append-only storage. Nothing is ever overwritten. Revisions to fundamentals or flows are new rows with a , preserving a full audit trail of what was knowable, and when. - Honest evaluation, not vanity metrics. Accuracy and precision are deliberately ignored. Sentinel evaluates with Information Coefficient, Spearman correlation, quintile spread, Sharpe ratio, turnover, and max drawdown — under pessimistic, realistic transaction-cost assumptions. - Adversarial self-testing. A dedicated test suite deliberately attacks the system — label shuffling, dead-feature injection, simulated future-leak attacks — to prove the validation pipeline actually catches what it claims to catch. - Earned complexity. No knowledge graphs, agents, or distributed infrastructure until the simpler system proves there's real signal to justify them. Currently running on SQLite by design. Current Status Phase 0 is architecturally and operationally complete. The system is in a 30-day live observation window — predicting relative outperformance for Indian banking equities, logging every prediction immutably, with zero new complexity permitted until the soak period concludes and the data has a chance to disagree with the model. Tech Stack Python · LightGBM · SQLite (deliberately, for now) · pandas · NSE/BSE data feeds · FinBERT Golden Rule Before adding any architecture: has reality earned this? If not, it doesn't get built yet.