Alternative data vendors sell narrative. Your job is to prove incremental alpha after costs, latency, and coverage bias. The papers below cover due diligence frameworks, empirical evidence on alt data value, and how quickly signals decay once discovered.

Industry and survey literature

Academic evidence on signal decay

Many alt data signals are repackaged price, volume, or known factors. These papers help you test whether a dataset adds incremental information.

101 Formulaic Alphas

Shows how crowded simple signals are. If your alt data factor correlates with these, edge is likely small.

arXiv

Text and news data

Text alt data is everywhere, but dictionary sentiment scores often miss return-relevant signal. Ke, Kelly, and Xiu show how supervised text mining can extract predictive information from news flow.

Predicting Returns With Text Data

Supervised text-mining pipeline for Dow Jones Newswires: predictive screening, topic-model word weights, and penalized aggregation into article-level scores. Benchmark for testing whether a vendor sentiment feed adds incremental alpha.

NBER

Due diligence checklist

  • Survivorship in panel construction: are delisted names included?
  • Timestamp: when was the data knowable to the market?
  • Coverage bias: does the vendor only cover large, liquid names?
  • All-in cost at your target universe size and rebalance frequency
  • Null test: shuffled labels should destroy predictive power

For new datasets, cross-check against open anomaly data at Open Asset Pricing (Chen-Zimmermann) before committing budget.