Every business question we get eventually turns into the same shape: given what happened, what happens next? Demand next quarter, churn next month, electricity load tomorrow at 6 pm. General-purpose language models are remarkable at explaining a forecast, but they are not built to compute one from a long numeric history.
So we trained our own family of prediction transformers. PE-1 slices a time series into patches, short contiguous windows of observations, and treats each patch as a token. A decoder-only transformer reads them in order and predicts the next stretch of the future, exactly the way a language model predicts the next word.
What that changes in practice
- Zero-shot forecasts. Point PE-1 at a series it has never seen and it forecasts on the spot, with accuracy that usually requires a bespoke model. Fine-tuning on your data goes further still.
- Honest error bars. Every prediction is a distribution, not a lone number: a central estimate and a calibrated band from P10 to P90, so you can weigh the risk.
- Any horizon, any cadence. Minutes to years, short histories to long ones, in a single efficient pass.
- Lean to run. Compact next to a general LLM, light enough to stream an answer the moment you ask.
How the Oracle uses it
When you ask the Oracle a question, a language layer turns it into a forecasting task, PE-1 does the maths, and the answer comes back in plain language: the number, the distribution and the reasoning behind it.
The same engine powers the forecasting work we deliver for clients, from inventory to energy peaks. If you have a history and a decision that depends on it, we would like to see both.