Contiguous Patch Masking

Contiguous Patch Masking is a method used by AI models that forecast data changing over time (like sales or visitor numbers). Instead of predicting one step at a time, the model hides an entire chunk of the future as a single block and fills in that whole hidden block in one go.

Think of filling in blanks on a test. If you fill them in one by one from left to right, getting the first blank wrong can throw off your guess for the next one, and errors pile up as you go further along. Contiguous Patch Masking works differently: it hides several blanks at once and looks at the surrounding context to fill in all the answers at the same time.

This approach speeds things up because it doesn't need repeated computation rounds, and it also reduces the problem of earlier errors carrying over and accumulating into later predictions. It's a design approach used when a time-series forecasting model needs to output multiple future time points at once.

The article explains that "TimesFM-3, building on research applying the Contiguous Patch Masking technique, masks out the entire future segment and then fills in all predictions simultaneously in a single forward pass." Unlike the previous version, which generated patches one at a time and accumulated latency and compounding errors, this technique computes multiple time points at once.