The biggest problem with the entire approach was the sheer number of parameters that needed to be specified. There are no “right” ones and you will almost always overfit.
We combined books and research papers to create a knowledge base. We then used that to create a research program to check if they can be applied to Indian stocks/futures for profit. Our threshold for success was outperformance against naive momentum/trend strategies. RL implementations tend to be complicated and compute intensive. They least they can do is beat simple systems. Also, India is an expensive venue to trade so these tests included those costs as well.
Across every reward, cost, state, action, algorithm, universe, and venue we tried, no RL agent beat classical baselines net of costs out-of-sample.
The stock that was sold usually went up after. At 20-days, you were 1.5× more likely to miss a >5% rally than to avoid a >5% slide.
The stock that was bought instead did about the same. A drift-positive basket was replaced with another drift-positive basket. The churn bought nothing and you paid transaction costs on top.
Trailing stop losses might be psychological crutches. But with everything else in finance that is designed to ease the pain of ownership, you end up paying for it.
We are not big fans of using VIX for timing equities. Strategies that seem to work on US market data fall apart when applied to Indian equity indices. When Alpha Architect published their follow-on VIX and Trend Following results, we decided to test it on our own NIFTY large-, mid- and small-cap indices.
India VIX decides how much recent history to use:
When volatility is calm, rank the assets by their 10-month performance. This changes slowly and avoids reacting to short-term noise.
When volatility is elevated, use 3-month performance so the strategy can respond faster.
When volatility is very high, use only the latest month’s performance.
tl;dr: it kind of works – beats the NIFTY 50 after 25bps of costs post-COVID.
However, with an on-going 30% drawdown, it is not exactly self-recommending. Having said that, its pre-COVID results might be a better indicator of how the strategy manages to side-step large shocks.
It does a semi-decent job of switching between different market-cap indices.
The full-period metrics (2009 – 2026) are promising. Seems to hold its own against buy & hold – lower drawdowns with similar(ish) returns.
While 25bps of costs should cover transaction costs, the problem is surviving capital gains taxes require a much higher returns.
Details of the approach, descriptions of the strategies and code are on github.
Gong, Rui and Lynch, John and Ogden, Richard, Skewness Managed Portfolios (May 31, 2026, SSRN) argue that returns to many prominent cross-sectional anomalies are driven by a small number of extreme, positively skewed stock returns. So explicitly selecting stocks with high expected skewness for the long side can substantially improve anomaly performance (summary).
If you look at the pre-COVID window, Omega variants out-performed the basic ones and Omega + Skew ruled them all. Also, the Skew variants out-performed the corresponding basic ones. However, zooming in on the post-COVID performance, we see that the effect of the Skew adjustment has attenuated a bit.
If you include the COVID crash, the big picture argues in favor of Omega + Skew.
You can go ahead and mark this as another useful tool in your toolbox.
