A sector rotation that matched its benchmark, and why we shelved it

The premise was ordinary. India's sectoral indices move in long, visible waves, and a rules-based rotation that leans into the leading sectors should beat holding all of them equally. We built it, we tested it, and it did not beat holding all of them equally.

What the development window said

Two findings from the development period were worth keeping, independent of whether the strategy survived.

The first is that India's sector indices show one-month continuation, not reversal. That is the opposite of the cross-sectional reversal effect people import from US single-stock work, and it flips the sign on any naive contrarian overlay.

The second is that risk-adjusted momentum ranks better than raw momentum. Dividing trailing return by trailing volatility before ranking sectors improved selection consistently, which is the sort of small, boring result that tends to hold up.

What the holdout said

We hold one out-of-sample window per hypothesis and we spend it once. On that window the strategy produced roughly 10.5% CAGR at a Sharpe of about 0.68.

Equal-weighting the same sector universe produced roughly the same thing.

There is a version of this write-up where that gets framed as "matched the benchmark with lower turnover" and quietly promoted. It matched the benchmark while taking rotation risk, paying rotation costs, and requiring a monthly decision that equal weight does not require. Matching is not a reason to deploy. It is a reason to stop.

The overlay that flattered itself

The single largest improvement during development was a 200-day moving average overlay on the sector sleeve. It lifted every development-period statistic we looked at.

On the holdout it was the worst change in the entire specification.

This is the recurring shape of the failure: the component that most improves an in-sample number is frequently the component most fitted to that sample. We now treat "this one change helped a lot" as a warning rather than a result.

Status

Shelved. The holdout for this hypothesis is spent, so any revised version of this idea needs a genuinely new evaluation window, not a re-run against the same data.