Reconstructing a Paper's Data and Universe

Most attempts to replicate a published trading result don't fail because the strategy logic was misunderstood — they fail because the replicator used a different universe of securities, or a different data source, without realizing it. Getting the underlying data right is unglamorous, but it's the step where the majority of replication effort should actually go.

Start with the universe: exactly which securities does the paper trade? "All NYSE-listed stocks" is not enough detail — did they include stocks that later delisted, or only ones that survived to the present (survivorship bias baked into the replication before it even starts)? Did they use point-in-time index membership, or today's membership applied retroactively? A momentum strategy tested on today's S&P 500 constituents, projected back twenty years, silently excludes every stock that got kicked out of the index for underperforming — which flatters the result. Next, the data source itself: prices, especially for anything before the last decade or two, differ across vendors due to different adjustments for splits, dividends, and corporate actions. A paper's "returns" figure is only reproducible if the replicator adjusts the same way.

The practical approach is to write down, before touching any code, a precise specification: security universe (with delisting and membership rules), date range, data vendor and adjustment methodology, and the exact fields used (open, close, adjusted close, volume). Only once that specification matches what the paper describes — as closely as the paper's own disclosure allows — does it make sense to test whether the actual strategy logic reproduces the claimed result. Testing the logic first, on whatever data happens to be lying around, just produces a number that isn't comparable to anything.

This work also surfaces the gaps in a paper's own disclosure, which are common — many papers describe their universe in a single sentence and leave the delisting treatment, exact vendor, and adjustment convention to be inferred. Where the paper is genuinely silent, the right move is to try the most standard, least favorable assumption first (include delisted names, use point-in-time membership) rather than whichever assumption happens to make the replication easier to match.