Financial time-series foundation model

A financial time-series foundation model is a large model pre-trained on broad market price history and reused as a general forecaster — the predict layer of algorithmic-trading. Kronos is the founding example: instead of natural-language text, it operates on K-line / candlestick (OHLCV) series — open, high, low, close, volume, amount.

The recipe, retargeted from text LLMs

The method borrows the text-LLM playbook and points it at price data:

1. A tokenizer quantizes the continuous, multi-dimensional candlestick stream into discrete tokens (Kronos uses a hierarchical scheme).
2. A decoder-only autoregressive Transformer predicts the next bar, then the next — the same tokenize-then-autoregress pattern behind text generation. This is the lineage shared with the open-weight text-LLM world (../llm-providers-wiki open-weight-models), differing in modality, not mechanism.
3. Pre-training spans many markets (Kronos: 45+ global exchanges), and the model is engineered for the high noise that distinguishes financial series from language.

The sub-field and the name clash

Domain-specific time-series foundation models are an active area: Amazon Chronos, Google TimesFM, Salesforce Moirai. Kronos is the financial-specialized entrant — and note the Kronos vs Chronos near-collision, easy to confuse.

Where it sits

A forecaster is not a full trading system: it outputs a prediction, which a strategy layer (RL policy or hand-coded rules in banbot) must still turn into sized positions and an execution plan. Its accuracy claims, being self-reported, want independent out-of-sample validation before trust.

Related

kronos-financial-foundation-model · algorithmic-trading · reinforcement-learning-trading · synthesis