Overview
Kronos models OHLCVA market bars as hierarchical coarse-and-fine tokens, then generates conditional forecasts from finite context. The series follows that path through tokenization, inference, scaling, adaptation, and evaluation, showing what the evidence supports—and why portfolio, risk, and execution remain outside the model.
Learning outcomes
- Kronos treats each OHLCVA bar as a joint observation, not six independent targets.
- Its tokenizer learns a hierarchical coarse-and-fine code that balances reconstruction fidelity with prediction cost.
- Causal generation, finite context, normalization, and sampling settings materially shape each forecast.
- Reported benchmark gains do not establish live profitability; portfolio, risk, and execution remain separate layers.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why does Kronos use a tokenizer for financial K-lines?
Kronos uses a learned tokenizer to turn joint OHLCVA observations into hierarchical discrete tokens that an autoregressive model can predict. Reconstruction objectives tie the code to the original market record.
How does Kronos generate a forecast?
Kronos predicts a coarse subtoken and then a conditionally linked fine subtoken from causal context, appends each pair to a rolling window, and decodes the resulting path back into continuous values.
What determines how much history Kronos can use?
Kronos's usable history is determined by the checkpoint's context limit, bar frequency, and configured lookback; older tokens can be shifted out as generation advances. A longer input is not unlimited memory.
Do Kronos benchmark results prove trading profitability?
No. Kronos's reported results are conditional on selected tasks, data, periods, baselines, and portfolio assumptions; live use still requires separate validation plus portfolio, risk, cost, and execution rules.