What the Results Mean—and What They Do Not

Kronos's experiments answer several different questions, and the results become misleading when those questions are collapsed into one score. This chapter interprets the reported forecasting, volatility, synthetic-data, and investment-simulation evidence by separating metric from meaning. The practical judgment is not whether Kronos “wins” in the abstract. It is which design choices the experiments support, how far those results travel beyond the tested protocol, and why a strong model benchmark is not the same thing as a live-trading guarantee.

Forecast metrics measure relationships, not wealth

The paper evaluates Kronos on five task families: price-series forecasting, return forecasting, realized-volatility forecasting, synthetic K-line generation, and investment simulation. The first three are predictive tasks, but they do not measure the same target. Price-series forecasting asks whether the predicted price channels move with the realized channels. Return forecasting asks about a derived price change. Volatility forecasting asks whether a measure of future variability is estimated accurately. Investment simulation adds a portfolio process around a model-derived signal.

For price and return forecasting, the paper reports Information Coefficient, or IC, and Rank Information Coefficient, or RankIC. IC measures the association between predicted and realized values; RankIC measures the association after comparing their order or rank. The distinction matters in finance. A model may get the ordering of opportunities mostly right while being poorly calibrated in magnitude, or it may match magnitudes without ranking assets usefully. Neither metric is a portfolio return, and neither specifies how a signal is converted into positions.

The paper's headline result is a large relative improvement in price-forecasting RankIC compared with selected baselines: it reports a 93% improvement over the strongest time-series foundation-model baseline and an 87% improvement over the best non-pre-trained baseline. Those numbers are meaningful only with their denominators. They describe relative performance under the paper's chosen assets, frequencies, horizons, dates, baselines, and evaluation procedure. They do not mean that every forecast is 93% more accurate, that error is reduced by 93%, or that an investor should expect a corresponding return.

The evaluation also compares Kronos against different kinds of models. Full-shot time-series models are trained on the downstream task. Zero-shot foundation models are evaluated without the same type of task-specific training. Econometric volatility models are specialized for a narrower target. Generative time-series models are designed for synthetic-data quality. These groups answer different comparison questions. A result against a general zero-shot model does not establish superiority over every trained financial strategy, and a result against a volatility model does not measure the same capability as a result against a price forecaster.

The ablation study is narrower and, in one sense, more informative. It compares continuous-space alternatives with discrete Kronos variants while attempting to keep model scale comparable. A direct autoregressive baseline predicts continuous values with a mean-squared-error objective. A probabilistic continuous baseline predicts a distribution with a negative-log-likelihood objective.