Confidential

Is there a stock-picking edge on Thailand’s SET that survives real costs, and still works today?

Sector: Quantitative research · Year: 2026

Summary

Research for a Thailand-based hedge fund looking at its home market, the Stock Exchange of Thailand. We built a clean history of every SET stock, including companies that later disappeared, and tested ten well-known stock-picking effects net of real trading costs, with the most recent three years locked away. One effect came close: buying stocks near their 52-week high. We tested it once on the locked years, and it did not hold up.

Challenge

Thailand raises two problems India did not. There is no free official archive of daily prices that includes delisted companies, and the exchange’s own subscription service goes back only five years, too short to tell a real effect from luck. Costs are also unusual. They come less from fees than from the minimum price step: on a typical liquid stock, one step is around 0.6% of the price, so a round trip costs about 1% once commission, exchange fees and VAT are added. The fund wanted to know whether any stock-picking effect survives that, and whether it still works in today’s market, before committing capital.

Approach

  • Built a survivorship-free history from a free daily archive of every security traded on the exchange since the 1970s, including companies later delisted. Separated 1,001 common stocks from funds, warrants and index rows, and traced 71 ticker changes so that a renamed company is not mistaken for a delisting.
  • Rebuilt splits, par changes, rights issues and dividends from a second, independent source, and checked the result against it: daily returns matched on 98.3% of days from 2006. Two days missing from the archive were documented rather than filled in.
  • Fixed the rules before testing: the 150 most-traded stocks, rebuilt every quarter using only information available at the time; buy at the next day’s open; every result net of commission, exchange fees, VAT and the price-step cost, then checked again at double the slippage. Costs this high rule out fast trading, so only quarterly rebalancing was tested.
  • Tested on 2006 to mid-2023. Everything after that was locked away as a holdout.
  • Ran ten well-known effects: momentum, 52-week high, low volatility, low beta, reversal, illiquidity and others. None passed. Buying stocks near their 52-week high came closest, beating the average of the 150 stocks by 5.1% a year, just short of the bar. It held up under every robustness check, including dropping all later-delisted companies and testing only from 2010.
  • Had the result reviewed independently before going further. The review tightened the start of the test to 2006, the point from which the data meets our quality standard, and removed a second effect whose pass had come down to rounding. Then we wrote down the pass rule and opened the locked years once, for the 52-week-high effect only.

Outcome

It did not hold up. Over the locked 2023–2026 period, the stocks near their 52-week highs fell 4.4% while the average of the 150 fell 15.2%. That margin was too small and too erratic to distinguish from luck (a t-statistic of 0.46 against the 1.0 required), and against the SET50 index it lost 20%. The test also exposed a clear weakness: the strategy lags badly when the market rebounds sharply. Its worst quarters (April–June 2009 and April–June 2020) were both strong recoveries, and the same pattern showed up again in 2025 and 2026. Our advice to the fund: no SET stock-picking effect tested here is ready for capital. With benchmarks this weak, a wider AI-driven idea search is unlikely to find more. The one avenue still untested is foreign and NVDR investor flows, which need their own study.

Stack

Python, pandas, NumPy

Published papers

  • Same Test, Two Markets — Ten textbook stock-picking effects, tested the same strict way on India’s NSE and Thailand’s SET. One came close. It faded on the years it had never seen. 9 pages, PDF.

More work