← Naman GuptaAll work
Case study · Quant research · 2023

35 alphas across two markets

Quant Researcher · 2023~10 min read
35+ signals shippedSharpe 1.25–2.91Top 0.5% global · rank 26Cross-corr < 0.7

TODO: Open with the headline finding — which signal class generalized across regions and which didn't. Then in one sentence: what surprised you. Then in one sentence: the bet a fund could actually make from this.

Context

TODO: What is an alpha, what is a contest like AlphaVerse, and what does "Sharpe 2.91" mean to a reader who has never traded? Two paragraphs max. Anchor the rest in why out-of-sample survival is the only metric that matters.

The setup

TODO: Universes (US large-cap, China A-shares), data vendors (WorldQuant BRAIN), evaluation regime (in-sample vs out-of-sample, holding periods, turnover caps, the "fitness" function that judges your work).

What I tried

TODO: Group the 35+ signals into 3–4 thematic buckets — e.g., momentum, mean-reversion, microstructure, cross-sectional anomalies. For each bucket: one example signal you can describe without revealing competitive IP, what it tried to capture, what it actually captured.

What generalized

TODO: Which 5–10 signals survived US → China and why. Hypothesis: structural anomalies (e.g., cross-listing arbitrage, calendar effects) travel; behavioural anomalies (e.g., retail momentum chasing) don't.

What broke out-of-sample

TODO: The signals that looked great in-sample and died on holdout. The honest reason — overfitting to specific volatility regimes, look-ahead via data alignment, survivor bias in the universe definition.

Numbers

| Metric | Best | Median (top 10) | |---|---|---| | Sharpe (in-sample) | TODO | TODO | | Sharpe (out-of-sample) | 2.91 | TODO | | Turnover | TODO | TODO | | Cross-correlation (with portfolio) | < 0.7 | TODO |

TODO: One paragraph on which row matters. Out-of-sample Sharpe is the only one a real fund would trust.

What I'd revisit

TODO: Two honest critiques. Candidates: the universe was static so survivor bias is structural; turnover caps were too generous; I didn't separately track signal decay (an alpha that works for 6 weeks is not the same as one that works for 6 quarters).

What this taught me about systems

TODO: Why this informs how you think about backend correctness too — "looks great in dev, dies in prod" has the same shape as "looks great in-sample, dies out-of-sample." The discipline transfers.