Forecasting AI models are pretrained on huge stretches of historical data, then usually need a second round of tweaking before they're trustworthy in production. A new paper takes stock of that second round and tries to make sense of it.
The paper looks at time series foundation models, or TSFMs, AI systems trained to spot patterns across many kinds of sequential data, from stock prices to server load to weather readings, and then reused for new forecasting tasks. Pretraining alone leaves gaps: models drift when deployed on data unlike their training set, struggle with tasks outside their original scope, and often have too little labeled data to adapt cleanly. Researchers have been patching these gaps with a scattered mix of post-training methods. The paper sorts that mix into five buckets: adjusting model parameters, feeding in better context, combining multiple models, cleaning up outputs and confidence estimates, and shrinking models for deployment.
The real value here isn't a new model, it's a map. Time series forecasting is quietly load-bearing for finance, logistics, and infrastructure monitoring, and right now every team adapting a pretrained model for those jobs is reinventing its own fix. A shared taxonomy makes it easier to compare methods and spot which gaps, like knowing how confident a forecast actually is, remain unsolved.
That gap in confidence estimation is the tell: forecasting tools are often judged on how right they look, not on how well they know when they're wrong.