KWL
← Field Notes

Foundations · Field Note 0003 · Apr 2026 · 5 min · KWL Research

We backtested our own product four times. Every test failed.

We pre-registered whether music predicts demand, tested it four ways, and it failed four times. We published every result, including the failures. Negative results are still results.

The result
4pre-registered tests
4failed
0survived
Publishedevery result, with the code

The question

We started with a tempting idea: that the music a market plays might predict where its spending is heading. It was not a wild guess. Peer-reviewed work in the Journal of Financial Economics shows the valence of a country’s music tracks its sentiment, and even its stock returns. If mood is revealed in music, perhaps demand follows.

It was worth testing properly. So we did.

The four tests

We pre-registered each test before running it, so we could not move the goalposts once we had seen the data.

Four pre-registered tests, four results
TestWhat it askedResult
Country forecastDoes national mood predict spending?Failed
Brand & category backtestDoes it predict brand or category demand?Failed
Powered nowcastFrance and Germany, properly poweredFailed
Domestic-only re-testA cleaner home-grown signalFailed

All four failed. On the broader outcome, adding music actually made the forecast worse than leaving it out.

What we did about it

We published every result, with the pre-registration and the code, failures included. The claim was falsifiable, and we reported the nulls, not just the wins.

We changed the product, not the data.

The science supports music as a contemporaneous mood proxy. It does not support it as a forecast. So that is exactly how we use it: a read of where a market’s mood and attention are now, not a prediction of demand.

What the tests left standing
What we dropped
Music as a contemporaneous mood read
Music as a demand forecast
A descriptive product
A predictive product
Pre-registered, published nulls
Confident, untested claims

Why publish the failures

An admitted gap earns more confidence than a confident guess.

In a category built on confident guesses, the working is the trust. Every figure in a Cadence read names its public source and date. Negative results are still results, and a predictive tier would reopen only on a future pre-registered test that actually clears.

References

  1. Edmans, A., Fernández-Pérez, A., Garel, A., & Indriawan, I. (2022). Music sentiment and stock returns around the world. Journal of Financial Economics, 145(2).
  2. Pre-registrations, test specifications and code for all four backtests are published with the Cadence methodology.