Data & drivers
Time series data quality: the clock is the data
Most forecast failures are not the model. They are a DST shift, a file of zeros that looked like a miracle, sales used as demand after stockouts, or a timestamp that is actually an invoice date. Fix the clock before you race architectures.
Updated Aug 26, 2026·8 min read
The clock is the data
Most forecast failures are not the architecture. They are a DST skip, a timestamp that is actually an invoice date, duplicated stamps after a vendor migration, or a file of zeros that looked like a demand miracle. Fix the clock before you race models. A tournament will happily crown a champion on a lie.
The clock is the data
Zeros that are not demand
Closed stores, offline meters, a clinic that does not run Sundays: those are missing, not zero interest. If you train on them as demand, you will forecast a ghost. Mark closures as calendar. Impute only with a rule you can run at serving — forward-fill a sensor, never invent customers.
Stockouts and delayed files
Sales during a stockout are censored demand. You observed the shelf, not the customer. Late files padded with zeros look like a miracle until the late rows land and rewrite history. Both will poison a live goal. Flag them. Prefer a demand proxy if you have one (traffic, quotes, failed scans).
- Profile gaps, duplicates, and sudden scale changes (units vs cases) before training.
- Do not silently resample time zones. Pick one clock and convert in the open.
- A pipe that died should look like an anomaly, not like a new season.
Profile first, then train
Connect a workspace and Predict.ai profiles the series — gaps, outliers, season — before a tournament spends credits. Anomalies on a live goal also catch a pipe that died, not only a world that moved. That is cheaper than finding out in S&OP that March was a file format.
You still own the semantics: what a zero means, what a stockout flag is. The product will not guess that a factory was closed. It will show you the hole.
FAQ
- Should I impute missing days?
- Only with a rule you can run at serving. Forward-fill a sensor. Do not invent demand on a day the store was closed. Mark closures as a calendar, not as zeros that look like no customers.
- What is censored demand?
- You observed sales, not demand, because the shelf was empty. Training on that history teaches the model to repeat the shortage. Flag stockouts or use a demand proxy if you have one.
Keep going
Guide
Why forecasts go stale in production
A forecast starts dying the moment the world moves and the model does not. New promotions, a competitor, a warehouse fire, a quiet change in how you record demand — live systems need a refresh cadence, drift checks, and a way to re-score the champion.
Guide
Anomaly detection for business metrics
A fixed threshold treats every Tuesday like every Saturday. A forecast band asks whether today is weird given what the series usually does at this clock. That is the difference between an on-call fire drill and a useful alert.
Guide
What is demand forecasting?
Demand forecasting estimates how much of a product or service people will want in a coming period, so you can buy, staff, and price before the rush — not after it has already walked in.
Use case
Retail demand forecasting
Anticipate demand by product, store, channel, and region before buying or allocation decisions are locked.
Use case
Energy load forecasting
Forecast demand by interval, feeder, zone, and customer class with weather-driven uncertainty.