How it works
Stock apps tell you what a stock did. This one tests what you think, and shows exactly how often its own models are wrong.
1. Prices in
Daily closes for the Nifty 50 and the index since 2014, pulled from Yahoo Finance by a Python pipeline. End of day, not live.
2. Models trained
LightGBM quantile regression for the 20-day return range. GARCH(1,1) and an LSTM (PyTorch) competing on 20-day volatility. A 3-state hidden Markov model for the market regime.
3. Every model put on trial
Walk-forward by calendar year: train on everything before, test on a year it never saw, and leave a 25-day gap so no answer leaks in. Each model is scored against a baseline a spreadsheet could compute.
4. Losers benched
If a model doesn't beat its baseline, the app uses the baseline instead and the scoreboard says why. Direction calls lost, so the app never makes them.
5. Your thesis tested
Each sentence is matched to the kind of claim it makes (a trend, volatility, the market backdrop, a business fundamental) and checked against the signal that can actually test it.
6. Tripwires written
For every sentence, the specific thing that would prove it wrong: a price, a percentile, a regime change. That's the part worth rereading next quarter.
In progress
- Three agents:a Researcher (sourced news and filings), a Model Builder (a formula-driven Excel valuation) and an Earnings Reviewer that reads each concall against your thesis. Each one cites a source and date for every claim, or says “could not verify”.
- Call language drift:FinBERT on BSE concall transcripts, joining the evidence only after a backtest.
Analysis only, never investment advice. The app does not say buy, sell or hold, and does not give price targets. Built as a portfolio project by Aryan.