K-lines get their own foundation model: pretrained on 45-plus exchanges, open weights forecast the next 24 hours
Synopsis
Kronos is the first open-source foundation model built for financial candlestick (K-line) sequences, pretrained on data from more than 45 global exchanges. General time-series foundation models are not designed for the high-noise character of financial data, so Kronos first quantizes continuous multi-dimensional K-lines such as open, high, low, close and volume into hierarchical discrete tokens, then pretrains an autoregressive Transformer on those tokens as one model for diverse quantitative tasks. The release includes pretrained weights of several sizes, fine-tuning scripts and a live demo forecasting BTC/USDT over the next 24 hours.
Interpretation
Financial K-line sequences now have a purpose-pretrained open foundation model instead of relying on general time-series models applied directly to market data. General time-series foundation models were not designed for financial data, while K-line sequences carry a distinctive high-noise character. Pretraining draws on data from more than 45 global exchanges, and a family of pretrained models with varying capacities is released.
The model turns continuous multi-dimensional K-lines into discrete tokens, then generates future market data token by token like a language model. A specialized tokenizer quantizes continuous multi-dimensional data such as open, high, low, close and volume into hierarchical discrete tokens, replacing direct modeling of continuous values. A large autoregressive Transformer is pretrained on these tokens and serves as a unified model for diverse quantitative tasks.
Raw K-lines become forecasts in a few lines of code, with preprocessing, normalization and inverse normalization handled by the prediction interface. Usage is wrapped in the KronosPredictor class, which takes historical K-lines, historical timestamps and future timestamps as its three main inputs. Kronos-small and Kronos-base have a maximum context of 512, and the predictor truncates longer lookbacks automatically; the live demo shows a BTC/USDT forecast for the next 24 hours.
Perspective
The demonstrated scope is an open foundation model for K-line sequences: pretraining data comes from more than 45 global exchanges, weights of varying capacity are released on the Hugging Face Hub, and fine-tuning scripts and a live demo accompany them. The demo page shows a BTC/USDT forecast for the next 24 hours, and Kronos-small and Kronos-base have a maximum context of 512, with the predictor truncating longer lookbacks automatically. It suits researchers and engineers working on quantitative modeling and financial time series.
The summary gives no comparative metrics on quantitative tasks, so the model's forecasting performance relative to general time-series foundation models has no numerical support here. Kronos-small and Kronos-base cap context at 512, so longer lookbacks are truncated, and the summary does not say how that limit affects long-history dependence. The live demo covers only the BTC/USDT next-24-hour setting.