What Data Would You Use?
Some interviews ask an open question instead of a puzzle. What data would you use to predict tomorrow's volume, to price a new contract, or to decide whether a signal is real? There is no single right answer. The interviewer listens for a structure, and for two common mistakes: naming data that would not exist at the moment of the decision, and never saying how the idea would be tested.
A frame in five steps
- Define the target. Say exactly what is predicted, over which horizon, and in which unit. "The price" is not a target. "Tomorrow's opening auction price, as a return on today's close" is.
- Say when the decision is made. Every input must be known at that moment. This one check rules out look-ahead.
- List data from closest to furthest. Start with the instrument itself, then closely related instruments, then news and fundamentals, and only then alternative data.
- Name the problems with each source. Frequency, gaps, survivorship, cost, and how far back the history goes.
- Say how you would test it. Choose a simple baseline, measure the error on data the model was not fitted to, and say what result would make you drop the idea.
An interviewer asks what data you would use to predict tomorrow's opening price of a stock listed in Amsterdam.
Target and timing. The opening auction price, as a return on today's close, predicted just before the auction ends.
Closest data. The stock's own close and its recent overnight gaps. The indicative opening price the exchange publishes while orders build up in the auction is the most direct evidence there is.
Related markets overnight. Index futures that trade before the cash market opens, the stock's US listing if it has one, the same sector in the US and Asian sessions, and the currency.
News. Company results, analyst changes and economic releases published before the open.
Test. The baseline is "the open equals the close". A model is useful only if it beats that baseline on days it was not fitted to.
The look-ahead trap
Data that describes the period you are predicting is not available when you predict it. Today's closing volume cannot predict today's open. A list of index members taken today cannot drive a test over the last ten years, because the companies that left the index are missing and the survivors make any strategy look better. The first mistake is look-ahead bias and the second is survivorship bias; the validation course treats both in depth.
Baselines first
A model without a baseline has no meaning. For a price, the baseline is usually "no change". For a rate or a volume, it is the recent average. State the baseline before any model, because it also tells the interviewer what "good" means in your answer. The same discipline decides whether a trading signal is real, which statistical edge covers.
Define the target and the decision time first. Then list data from closest to furthest, check that each input is known at the decision time, and name the baseline the idea must beat.
Say the simple sources before the exotic ones. Satellite images and card spending make a fine last line, but an answer that starts with them suggests the basics were skipped.