Confidential

Is there a stock-picking edge on India’s NSE that survives real trading costs?

Sector: Quantitative research · Year: 2026

Summary

Research for a Thailand-based hedge fund looking to trade India’s National Stock Exchange. We built a clean history of every NSE stock since 2006, tested ten well-known stock-picking effects, then had an AI model propose 306 new ideas and tested each one the same strict way. Nothing cleared the bar. One weak but real effect did show up: stocks where most of the traded volume is taken as delivery tend to do a little better.

Challenge

Stock-picking ideas are easy to make look good on paper and hard to make work with real money. India adds two specific problems. Trading is expensive: securities transaction tax alone is 0.1% each way, so buying and selling costs about 0.4% before anything else goes wrong. And most free price histories only contain companies that still exist today, which quietly flatters every backtest. The fund needed to know whether any edge survives both, before committing capital.

Approach

  • Built a survivorship-free price history from the NSE’s own daily files: every stock that traded, including companies that were later delisted, suspended or merged, with splits and bonuses adjusted from the exchange’s corporate-action records.
  • Fixed the rules before testing: the 500 most-traded stocks, rebuilt every six months using only information available at the time; buy at the next day’s open, never at a price already seen; every result net of tax, stamp duty, exchange charges and slippage, about 0.4% per round trip, and checked again at about 0.6%.
  • Tested on 2007 to mid-2023 only. Everything after that was locked away as a holdout and never looked at, so it is still available as a clean final test.
  • Ran ten well-known effects first as the benchmark: momentum, 52-week high, low volatility, low beta, short-term reversal, illiquidity and others. None passed. The best, momentum, beat the average of the 500 stocks by 5.2% a year, but not convincingly enough over the period.
  • Had an AI model propose 306 new ideas, using the exchange’s daily delivery data, corporate events and surveillance listings as well as prices. Each idea went through the same backtest, a check for any use of future information, and a deflated Sharpe ratio, a test that asks how good a result has to be before you believe it, given how many ideas were tried.
  • Caught and removed ideas that were secretly using the future. The strongest-looking one was ranking stocks by price levels that had already been adjusted for splits that happened years later. Measured properly, its result roughly halved. The check now rejects that pattern automatically; 63 ideas were removed this way.

Outcome

No idea cleared the bar. With 192 genuinely different ideas tested, a result needs a risk-adjusted excess return (Sharpe ratio) of about 1.1 to be believable. The best candidate reached 0.72: stocks with persistently high delivery, which beat the average of the 500 stocks by 7.1% a year. A second, independent idea built on the same delivery signal near 52-week highs reached 0.65 at 3.2% a year. The effect is probably real, people holding shares rather than day-trading them, but it is roughly two-thirds of the way to the bar, and that gap is the gap between interesting and bankable. Every idea that traded often was wiped out by costs. Our advice to the fund: no NSE stock-picking edge found here survives realistic Indian costs and the number of ideas tested. If the delivery effect is worth pursuing, it needs its own study with its own rules, a single look at the locked holdout, and months of paper trading, with the warning that even if it is real, it may not pay after costs.

Stack

Python, pandas, SQLite, Gemini

More work