Music Equities

When a quarter closes, every show and every pageview is already in the public record, but the revenue number is not published for roughly another 35 days. The idea was to read that past faster than the market does: nowcast at quarter-end plus 10 trading days, which is on average 22 days before the print. Nothing here predicts the future.

The thesis behind it: small-cap music equities are under-modelled, and alternative data might reach them where the big quants don't bother.

All free and point-in-time. SEC XBRL filings for revenue and deferred revenue, taking only the first-filed value: 287 company-quarters, with 64 restatements caught. Real earnings dates from 8-K filings (254 of 287 from actual filings). Monthly Wikipedia pageviews for 24 curated entities, 133 months across 36 page titles. Adjusted total-return prices over 2,765 trading days.

The universe was nine US-listed companies, frozen before the first run, including Sonos as a deliberate placebo: music attention should have no power over a speaker company's revenue. Final panel: 246 company-quarters, 173 complete, 119 for development and 54 sealed in a lockbox.

The plan was written and frozen before any model ran. Stage A: nowcast revenue growth in an anchored walk-forward (test years 2019 to 2023, purge plus a 21-trading-day embargo at every seam) and compare it with three naive baselines. Stage B: trade the surprise, charging each name its own liquidity-tier cost of 10, 25 or 50 basis points per side.

The bars, set in advance: beat all three baselines; beat a shuffled-label null at p < 0.05; show positive skill in at least 3 of 5 folds; show materially less skill on the Sonos placebo; and, for the trade, a Sharpe above 0.5 with a confidence interval excluding zero. COVID years 2020–21 were excluded from the headline in writing before anything ran, and everything from 2024-07-01 onward stayed in a sealed lockbox.

To check the harness itself: Gaussian-noise features, shuffled labels (200 draws), random Wikipedia pages in place of the curated map, and an extra quarter of lag.

Every pre-registered bar failed except the placebo. The nowcast beat 1 of 3 baselines. Against the shuffled-label null it sat at p = 0.68. Skill was positive in 0 of 5 folds. The trade's Sharpe was −0.66, with a confidence interval of [−1.68, 0.51].

The most damning number: an attention series built from the Wikipedia pages for Photosynthesis, Chess and Coffee scored −0.78, better than the hand-curated music map's −2.36. And in an encompassing test, persistence alone had an error (MAE) of 0.2032; adding attention raised it to 0.3435. Attention made the forecast worse.

The harness was clean: 0% of 200 Gaussian-noise draws earned positive skill. One trap was avoided only because of the plan: over all development years the trade showed a Sharpe of +0.27, almost entirely from 2020, when it returned +119.9% by being long COVID-recovery names. Those years had been excluded before any model ran.

Change the unit of observation. Modelling per-show economics instead of per-quarter revenue turns roughly 200 observations into roughly 50,000. That needs a setlist.fm key and Ticketmaster Discovery data for forward ticket inventory, which can never be backfilled, so the nightly collector should start regardless.

Then widen the sample: more segment history and more companies with genuinely forward-sold inventory, including non-music venue and ticketing operators. Only once the development estimate is credible on its own: one confirmatory run on the lockbox. Not before.