An overlay that reversed sign out of sample
A rotation overlay on NSE strategy indices looked additive in development and inverted on the holdout. The third time this exact failure has shown up.
- momentum
- risk
- null result
NSE publishes a family of strategy indices: low volatility, quality, alpha, value, momentum, and several blends. Each is a rules-based portfolio with a public methodology. If those rules capture real and distinct exposures, then rotating between them ought to be tractable, because the underlying objects are transparent in a way that individual stocks are not.
We built a barbell overlay that tilted between defensive and aggressive index sleeves based on a regime read.
The number that mattered
In development the overlay produced an information ratio of about +0.39 against the static barbell.
On the holdout it produced about -0.43.
Figure 1
Overlay information ratio against the static barbell, by evaluation window.
| Window | Information ratio |
|---|---|
| Development | +0.39 |
| Holdout | -0.43 |
Research output on historical data, shown for illustration only. Not audited, not a track record, and not indicative of future results.
Selection information ratio across the full period came out at roughly +0.02, which is another way of saying the overlay did nothing except move the outcome around.
A sign flip of that size is more informative than a small positive would have been. An overlay that adds a little is hard to distinguish from noise. An overlay that reverses is telling you the development-period relationship was the artefact.
Why we believe this rather than re-tuning
This is the third time a timing overlay on Indian equity factor exposure has failed for us in the same way. Different universes, different regime definitions, different research streams, same shape: a positive development-period number that inverts out of sample.
At some point the correct read is not "the parameterisation was wrong." It is that the regime signal does not carry the information the overlay needs, and each additional attempt is another trial against the same underlying data. We count failed searches into the multiple-testing budget for exactly this reason, so the bar for the fourth attempt is higher than it was for the first.
What survives
The static barbell survives. Holding a fixed mix of defensive and aggressive sleeves and rebalancing on a calendar did fine, without a regime call.
That is the useful residue of this work. Not every research programme ends with a new signal. This one ended with a clearer sense of which decision in the stack was not worth making.
Status
Do not redeploy the timing overlay. The static allocation stands.
This note describes research on historical data and is published for illustration only. Figures are not audited, are not a track record, and are not indicative of future results. Nothing here is an offer, a solicitation, investment advice, or a recommendation to transact in any security.