Multibagger Lab: Blind Model-vs-Baseline Stock Test
A controlled experiment on 4,440 anonymised NSE stocks comparing a language model's picks against gradient boosting and momentum, with a hash-chained forward ledger.
Split- and ISIN-stitched price history from 2016, a blind backtest where Claude's top decile hit 12.2% versus a 6.9% base rate, a GBM baseline that did better, a weekly forward ledger that excludes circuit-locked stocks, and an 8-factor alpha study.
- Blind test on anonymised stocks with a stated base rate
- Model lift 1.76x vs GBM 2.06x, reported honestly
- Hash-chained forward ledger, graded on fixed dates
- Circuit-locked stocks excluded by rule
Can an AI pick stocks that double? Rather than argue, the lab measures it. NSE bhavcopy archives from 2016 were stitched through splits and ISIN changes, then 4,440 stocks were anonymised so the model saw only numbers. On the question of a two-times move, the base rate was 6.9 percent; the model's top decile achieved 12.2 percent (a 1.76x lift), a gradient-boosting baseline 14.2 percent, and simple momentum 11.5 percent. Five-times and ten-times samples were too small to score, and the model was weakest in the most recent two years. Those results are stated plainly.
A weekly scheduled task writes the current picks to a hash-chained forward ledger so nothing can be edited after the fact, with grading dates set months ahead. A rule from me is enforced in code: a stock locked at the upper circuit with no sellers is unbuyable, so it is excluded from picks and never graded.
A companion alpha study built an eight-factor monthly top-20 portfolio with liquidity and cost constraints; two factors that looked good in-sample did not hold out of sample and were dropped.
Research only; shared as engineering work, not financial advice.
How it works
- Bhavcopies and corporate actions are stitched into adjusted histories
- Stocks anonymised and features exported per year
- Model, GBM and momentum each rank the universe
- Hit rates computed against 2x/5x/10x outcomes
- Weekly forward picks appended to a hash-chained ledger
Software and automation are infrastructure, not financial advice. No performance guarantees. Every live system needs staged testing, risk limits, and owner approval.