A language model trained on internet-scale data through 2024 has already seen the outcomes of events that occurred before its cutoff. If you use it to “predict” stock movements in 2022, the model may retrieve the answer from its training data rather than reason from the available evidence. The prediction is contaminated by knowledge the predictor should not have.
DatedGPT addresses this by training twelve separate 1.3B-parameter models, each on temporally partitioned data with strict annual cutoffs from 2013 to 2024 (arXiv:2603.11838). The 2019 model has never seen data from 2020. Perplexity-based probing confirms that each model's knowledge is effectively bounded by its cutoff year.
The architectural intervention is brute force: don't filter the data; train a new model. This works because the contamination is not in any specific document but distributed across the entire pretraining corpus. You cannot remove knowledge of a specific future event from a model trained on data that collectively implies that event. The entanglement of future knowledge with general capability makes surgical removal impractical.
The cost is twelve models where one would suffice, each smaller than frontier models and therefore less capable. The tradeoff is explicit: uncontaminated prediction requires reduced capability. A model that genuinely doesn't know the future is a model that knows less about everything.
For financial backtesting, the tradeoff is worth it. A contaminated backtest tells you nothing — it measures recall, not prediction. The temporal fence ensures that evaluation measures the ability to reason from evidence, not the ability to remember outcomes.