Your data is not as bad as you think

Six years of messy spreadsheets beats a clean warehouse with four months of history, most of the time.

The most common reason companies delay is a belief that their data is not ready. Usually it is, and the belief is doing more damage than the data.

What matters for most forecasting and classification work is history and consistency of meaning, not tidiness of format. Six years of spreadsheets with inconsistent column names but a stable definition of what a sale is will produce a better model than a beautifully modelled warehouse containing four months of records. Formatting is a week of work. History cannot be bought.

The problems that genuinely stop a project are different, and worth naming. A field whose meaning changed silently two years ago. Records that only exist for successful outcomes, so the model never sees a failure. Human judgement that was never written down anywhere, only held in the head of the person who left. Those are real, and no amount of engineering recovers them.

That is what the week-two data check is for. It takes a few days, it costs a fraction of a build, and its most valuable output is occasionally the sentence: wait six months, start recording this one field, then call us.