The First Working Mental Model: From Market Bar to Predicted Path

Kronos can now be understood as one connected process rather than as a tokenizer followed by an unrelated language model. This checkpoint brings the established pieces together around a practical judgment: when Kronos produces a forecast, what exactly has been observed, compressed, predicted, and reconstructed? Keeping those states separate prevents two common mistakes—treating a token ID as if it were a price, and treating a decoded forecast as if it were already an investment decision.

One market bar, four different objects

Start with a single K-line record. It is a continuous observation containing open, high, low, close, volume, and amount for a chosen interval. At this point, the record belongs to the data domain. It has the scale, measurement conventions, and missing-information limits of the supplied market data. It may be a useful summary of the interval, but it is not yet a model token and it does not carry a prediction about what comes next.

The tokenizer maps that observation into a learned latent representation and then into a hierarchical discrete token. The token is not a compressed spelling of the original numbers in any human-readable sense. It is a code that the tokenizer's decoder knows how to interpret. Its first part is trained to support a coarser reconstruction; the complete pair supports a more detailed reconstruction. The token therefore belongs to the model's representation space, where the important question is not whether its integer value is large or small, but what continuous configurations the learned decoder associates with it.

A history of these tokens is a third object. It is no longer a table of raw market measurements, yet it is not a future forecast. It is the context from which the decoder-only Transformer estimates the next token distribution. The context contains the joint representation of prior bars and, as generation advances, may also contain tokens the model generated itself. Calendar embeddings add information about minute, hour, weekday, day, and month, but they do not turn the context into a complete account of news, order flow, or economic causes.

The generated token path is a fourth object. At each future step, the model samples a coarse subtoken, uses that choice to condition the fine prediction, and appends the pair to its rolling context. The result is a sequence of discrete decisions. It is not yet a price chart. The tokenizer must decode the sequence into continuous values, after which the public predictor restores the scale of the input window. The final DataFrame is therefore the endpoint of a chain of transformations, not the direct output of one numerical regression head.

Imagine an observed bar whose closing price is ordinary but whose range and trading activity are unusual. The tokenizer may encode that joint configuration differently from a quiet bar with the same close. The generator then sees the resulting token in the context of earlier bars and timestamps, rather than seeing “unusual” as a label written by a human.