TimesFM-ICF, or In-Context Fine-tuning, builds on the base TimesFM model by continuing pre-training with new context: the forecast history plus all in-context examples. The aim is to help the model separate the history it is forecasting from the examples it is meant to learn from.
TimesFM is described as a patched decoder that tokenizes every 32 contiguous timepoints as an input token and applies a transformer stack on top of the sequence of input tokens to generate output tokens. It then applies a shared multilayer perceptron (MLP) to translate each output token back to a time series of 128 timepoints.
In TimesFM-ICF, a special, learnable “common separator token” is added after each set of numbers. The source says this is intended to act like a digital “stop sign” or a “new paragraph” symbol, helping the model avoid confusing the forecasting history with in-context examples.
That separation matters because, without it, the model could merge different numeric sequences into one continuous stream of data and misread distinct patterns as a single trend. With the separators in place, the model can attend to a prior example without mixing it up with the data it is currently trying to predict.
The approach is designed to let the model learn from patterns in past examples and apply that knowledge to the current forecast. The source gives an example of the model learning that “all the store sales are showing consistent, directional trends lately, so I should predict an upward trend for my new store’s sunscreen sales.”
Source: research.google.
Companies can share verified announcements through Newz9’s international press release submission page.
