Your First Hour With a New Dataset

The temptation with a new dataset is to jump straight to building a signal. The first hour is better spent trying to break the data instead: checking its coverage (which tickers, which dates, any obvious gaps in the middle rather than just at the edges), its granularity (is a "daily" field actually stamped once a day, or is it a snapshot at an inconsistent time), and whether any rows are duplicated or a symbol appears twice under two different identifiers.

A specific trap worth checking early is whether the data was point-in-time or has been overwritten with hindsight — for example, a "market cap" field which uses today's shares outstanding applied retroactively to old prices, which silently inflates or deflates historical values versus what was actually known at the time. Another cheap check is plotting the number of rows per day over the full history; a sudden jump or drop usually marks a vendor change, a new listing venue being added, or a change in how the field was calculated, any of which can quietly break a backtest that spans the transition.

None of this requires sophisticated tooling — mostly counting rows, plotting coverage over time, and spot-checking a handful of known dates against a second source — but skipping it is the single most common reason a promising signal turns out to be a data artifact once someone else tries to reproduce it.