A Simple AI Took 3rd Place Predicting the Future — No Chatbots Needed
Better forecasts mean smarter stocking, staffing and energy bills — without AI hype.
Two researchers, Nathan Thierry and Andre-Louis Rochet, have released a forecasting system called TW3Cast that predicts future numbers — sales, electricity demand, traffic, that kind of thing — by learning from past patterns. On GIFT-Eval, a public scoreboard where 130 teams compete on how accurately their systems predict unseen data, TW3Cast landed in 3rd place. The two entries ahead of it both use AI agents, meaning software that reasons step by step with language models. TW3Cast uses neither. That's the surprising part: a plainer approach beat almost everything fancier.
The trick is a "frozen router." Think of it like a laminated cheat sheet taped to the wall. For each of 97 different forecasting situations — different datasets, time scales and how far ahead you need to look — the sheet says which model to use. Sometimes it's a specialist model tuned for that specific job; sometimes it's a blend of several. Every choice on that sheet was decided once, using only past training data, then locked in. Nothing is decided live, so results are fast and repeatable.
The system's building blocks are public "foundation models" — AI trained on huge amounts of data that can be lightly adjusted for specific tasks. Adjusting them is cheap: each new candidate specialist costs a few megabytes and minutes of computer time, and if it flops, nothing breaks. The team also built three safety checks to stop the system from fooling itself, including one that guards against models that may have already seen the test data. Served alone, the best single model ranked 33.8 on average; the full router hit 19.4, where lower is better.
Everything is released publicly — the routing table, the model list, the exact scores — and one script reproduces every number in the paper. That transparency is rare and useful. The catch: these are academic datasets, not your company's messy spreadsheets, and 3rd place is not 1st. Still, it's a signal that careful engineering, not just bigger AI, still wins.
- Time-series forecasting (predicting future numbers from past patterns) is the tech behind demand planning, staffing and energy use.
- TW3Cast ranked 3rd of 130 on a public accuracy scoreboard while using no AI agents or language models — just a fixed lookup table choosing among pre-trained models.
- It scored 19.4 on average versus 33.8 for the best single model and 38.0 for a one-size-fits-all approach — and the team published everything so anyone can verify it.
Why It Matters
Cheaper, simpler forecasting could mean less waste, better stock levels and lower bills for ordinary businesses.