Introducing Chronos: Learning the Language of Time Series for after-work quants

Takeaway. Do not invent a new TS architecture first. Scale a series by its mean absolute value, quantize into a fixed vocabulary, train a stock T5 with cross-entropy, then sample tokens and dequantize. That is Chronos: regression via classification, zero-shot on new series.

Chronos: Learning the Language of Time Series (Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, and co-authors, 2024; arXiv:2403.07815). AWS AI Labs (with university affiliates). Code / weights: amazon-science/chronos-forecasting.

Fit for learners: Tokenize returns like language and zero-shot forecast — fun Hugging Face-style playground for evening experiments.

This Hub note is a plain-language intro. Not investment advice.

The problem

Deep forecasters still train per dataset. Prompted LLMs (LLMTime) need huge models. Purpose-built foundation models add lags, patches, and real-valued heads. Public TS data is tiny vs text. Can a vanilla language model, with only a tokenizer change, become a general probabilistic forecaster?

The idea / core intuition

• Mean scaling: (s = \frac{1}{C}\sum |x_i|), (\tilde x = x / s) (keeps zeros). Uniform bins on ([-15, 15]), vocab 4096 including PAD/EOS. No calendar features.
• Loss: next-token categorical CE — not Gaussian CRPS. Nearby bins must be learned from data.
• T5 Mini→Large (20M–710M); also a GPT-2 check. Context 512, prediction 64 in training. 200k AdamW steps, batch 256.
• Data hunger: TSMixup (Dirichlet mix of (k\le 3) scaled series) + KernelSynth (random GP kernel sums/products). Train mix ~9:1 real-aug : synthetic. ~28 train datasets / 84B tokens-scale observations.
• 42-dataset protocol: Benchmark I (in-domain) Chronos-T5 Base/Large beat local stats and task-specific PatchTST/DeepAR on aggregated relative WQL/MASE. Benchmark II (zero-shot) sits with the best per-task deep models; a cheap 1k-step fine-tune of Small takes the top slot.
• Limits they show: exponential trends, overflow if spikes exceed (15s), precision loss if variance (\ll) mean, slower inference than a tiny TFT.

Hugging Face ChronosPipeline is the evening path — not training 710M from scratch.