The Data and Adaptation Loop Behind Generalization

Kronos's generalization claim begins before the Transformer sees a token. It depends on which market records enter the pre-training corpus, which records are removed or separated, and how a local dataset is used when the released checkpoint is adapted. This chapter follows that lifecycle from raw K-lines to a reusable checkpoint and then to a local fine-tuning experiment. The practical judgment is simple: a larger corpus or a successful training run is meaningful only when the data transitions and evaluation boundary are understood.

How raw market records become training material

The paper describes Kronos's pre-training corpus as a large, heterogeneous collection of financial K-line records across multiple asset classes, exchanges, and temporal granularities. The project README presents the headline as more than 12 billion records from over 45 global exchanges. The paper's appendix describes the dataset as aggregated from over 40 exchanges across more than 30 countries. Those phrases may reflect different rounding or scope choices, so the responsible conclusion is that the corpus is reported as multi-billion-record and multi-exchange financial data—not that the retained sources establish one exact, independently auditable exchange count.

The scale matters because foundation-model pre-training is a distributional claim. Kronos is meant to reuse patterns across instruments and frequencies, so it needs exposure broader than one local ticker or one short historical period. But breadth does not automatically equal quality. Financial data can contain missing values, structural breaks, illiquid intervals, stagnant prices, and feed errors. If these artifacts are treated as ordinary examples, the model may learn relationships that describe the data pipeline rather than the market behavior the project intends to model.

The paper therefore describes cleaning as a multi-stage process. Missing prices are treated as hard boundaries: instead of inventing a price and continuing through it, the series is divided into valid contiguous subsequences. Missing volume and amount are treated differently because the project regards them as auxiliary covariates; they are imputed with zero. The paper also reports randomly setting volume and amount to zero for 5% of training samples, encouraging the model to remain useful when those fields are sparse or unavailable.

These choices preserve different kinds of information. A missing price can change the path itself, so carrying a fabricated value across the gap would create artificial continuity. A missing auxiliary field can be represented as absent activity under the training convention, allowing the model to learn price-based behavior as well. Neither policy is a universal rule for all financial datasets. They are part of Kronos's data contract and must be reconsidered when a local source uses different missing-value semantics.

The pipeline also filters low-quality segments rather than treating every valid-looking row as suitable training material. It partitions series around large price discontinuities, using frequency-dependent thresholds to isolate events such as rollovers, splits, or distribution artifacts. Within the resulting segments, it screens for sustained illiquidity and prolonged price stagnation. Only subsequences that satisfy the relevant minimum-length requirement are retained.

The frequency dependence is important.