Skip to main content

Alt-Data Edge Decay & Crowding

Updated Jun 24, 2026 at 2:35pm

Research Draft Medium 1,312 words

Alternative data (credit-card panels, satellite imagery, web-scraped prices, geolocation foot-traffic, app downloads, shipping/AIS feeds) can deliver real predictive edge over consensus estimates — but that edge is perishable. Edge decay is the erosion of a dataset's incremental alpha as it diffuses from a few sophisticated buyers to many, and as the mispricing it reveals gets arbitraged into prices. Crowding is the structural consequence: when many funds trade the same signal, their positions become correlated, capacity shrinks, and they are exposed to synchronized unwinds. The core tension is that the very act of monetizing a signal — and especially commercializing or publishing it — accelerates its own demise. Alt-data alpha is therefore best understood as a depleting resource, not a permanent moat.

How decay and crowding form

Three distinct mechanisms degrade an alt-data edge, and they compound:

1. Diffusion / commoditization. A dataset starts proprietary or expensive, accessible to a handful of funds. As the vendor scales sales (more clients = more revenue), the same signal is priced in by more participants. Industry commentary commonly cites a window of roughly 12–24 months of meaningful alpha before a novel dataset becomes broadly adopted (navnoorbawa/Substack synthesis; treat as a rule of thumb, not a measured constant). 2. Price impact / arbitrage. Once enough capital trades a signal, their aggregate order flow pushes prices toward the signal's prediction before the rest of the market reacts, leaving less residual to capture. In the limit of many identical models, the opportunity is arbitraged at the moment it appears. 3. Publication / disclosure effect. Removing uncertainty about whether a signal works is itself alpha-destroying. The most rigorous evidence comes not from alt-data per se but from the broader factor literature: McLean & Pontiff (2016, Journal of Finance) studied 97 published predictors and found portfolio returns were 26% lower out-of-sample and 58% lower post-publication — implying roughly a 32% publication-attributable decline beyond ordinary statistical overfitting. The mechanism (sophisticated traders learn from disclosure) applies directly to alt-data once a working paper, vendor white paper, or conference talk demonstrates a signal's value.

Crowding then emerges as a positioning risk. Khandani & Lo (2011, documenting the August 2007 "quant quake") showed that funds whose returns are nearly uncorrelated in normal regimes can become almost perfectly correlated under forced deleveraging — everyone selling the same names to the same buyers at once. Crowded alt-data signals inherit this regime-dependent tail: capacity squeezes in calm markets, synchronized losses in stressed ones.

How it's used in practice

Serious alt-data shops treat decay as something to monitor and manage, not avoid:

  • Rolling signal validation / sunsetting. Signals are tracked on a rolling out-of-sample basis (e.g., decaying information coefficient, hit-rate, or factor return) and retired once they stop predicting. The pipeline is a treadmill: new datasets onboarded as old ones fade.
  • Speed and execution as the real moat. A recurring counter-thesis (e.g., Izydorczyk, Magis) is that for commoditized data like credit-card panels, the edge is not exclusivity but the infrastructure to clean, map, cross-reference, and trade it fast — entity resolution, panel-bias correction, mapping spend to tickers, and acting before the print. Decay hits the naive signal; well-engineered users persist.
  • Capacity and crowding controls. Estimating how much capital a signal can absorb before its own impact kills it; diversifying across many weakly-correlated datasets so no single crowded factor dominates; and watching crowding proxies (short-interest overlap, factor-return correlation, days-to-cover) to size down before unwinds.
  • First-mover monetization. Because the window is finite, the economics favor exploiting a novel dataset hard and early, then rotating — consistent with the observed pattern that the biggest returns accrue to the earliest, most-restricted users.

Adoption, debate & evidence

Decay is widely accepted in principle — vendors, allocators, and academics agree alt-data alpha is finite. The genuine disputes are about magnitude and inevitability:

  • "Alpha is dead on distribution" vs. "edge is in execution." One camp argues wider distribution mechanically kills returns; the opposing camp (and several vendors, who are not disinterested) argues commoditized data still pays because few buyers have the operational chops to extract it. The credit-card case is the canonical battleground — it is simultaneously "the definition of commoditized" and still demonstrably profitable for some funds. Both can be true: the easy signal decayed; the engineered signal did not.
  • Measured decay. The cleanest hard numbers are from the factor-predictability literature (McLean–Pontiff's 58%/26%), not alt-data specifically — a transfer-of-evidence caveat worth flagging. For alt-data directly, Katona, Painter, Patatoukas & Zeng (Evidence from Outer Space) documented a satellite parking-lot strategy worth ~4–5% over the three days around earnings (RS Metrics imagery, 44 major U.S. retailers, 2011–2017), and the authors explicitly expect this to be "competed away" as adoption grows — even noting their own paper may accelerate it. That is a vivid illustration but a single case study, not a population estimate.
  • Convergence / homogenization risk. Recent work argues that as many funds train similar models on the same datasets, signals converge, raising correlation and systemic fragility (the "AI-driven alpha decay" thesis; arXiv preprints — treat as early/contested). The cited preprint's empirical section reports simulated institutional portfolio convergence rising ~42% over its 2013–2024 sample (calibrated to SEC 13F holdings, not observed fund returns); its prediction that return dispersion falls with AI adoption is a theoretical corollary, not a measured statistic. These are modeled results and should be cited cautiously.

Folklore vs. measured: the precise "12–24 month half-life" is industry lore, not an audited statistic. The robust, peer-reviewed finding is qualitative-but-strong: published/disclosed predictors decay materially, faster for higher in-sample returns and in liquid, easy-to-arbitrage stocks.

Strengths & limitations

Recognizing decay is itself an edge: it disciplines you to validate out-of-sample, sunset stale signals, and avoid paying premium prices for already-crowded data. It works best as a framework — explaining why backtests overstate live returns and why a once-great signal quietly stops paying.

It fails as a precise forecasting tool. Decay rates are dataset-, regime-, and execution-specific; there is no reliable formula for "this signal has N months left." The #1 misuse is over-fitting to a freshly published or vendor-marketed signal and assuming the backtested Sharpe will persist — exactly the cohort McLean–Pontiff show decays most. A close second is ignoring crowding until a deleveraging event reveals that your "diversified" signals were the same trade.

Sources

Disputes flagged: the "12–24 month half-life" is industry rule-of-thumb, not measured; the hard decay numbers come from the factor literature (transfer-of-evidence to alt-data); the AI-homogenization convergence figures are simulated (13F-calibrated), not clean empirics, and the dispersion-decline claim is theoretical; vendor sources have a commercial interest in the "edge survives" position.