Kronos: Turning K-Line Data into Market Language

The Dilemma: Why General Time-Series Models Struggle in Financial Markets

In recent years, we have witnessed large language models (LLMs) like GPT learn human language, logic, and even creativity by “reading” massive text corpora. The success of the GPT paradigm has inspired the development of time-series foundation models such as TimeGPT. However, financial markets remain a highly challenging application scenario for these models. The core issue lies in the double mismatch between the “foundation of model training” and the “characteristics of financial data.”

On one hand, the pre-training corpora of existing general time-series foundation models primarily consist of physical-scenario data such as power load, traffic flow, and solar generation. Financial sequences constitute a negligible portion of these corpora. General models learn stable, physics-driven patterns—such as daily cycles (“high power usage during the day, low at night”) or traffic “rush hour” bimodal distributions—from this data. These patterns exhibit strong stationarity and predictability. In contrast, K-line (candlestick) sequences are characterized by low signal-to-noise ratios and strong non-stationarity. These traits clash severely with the inductive biases of general time-series foundation models, resulting in performance often inferior to simple linear models and an inability to generalize effectively across broad quantitative finance scenarios.

On the other hand, K-lines are multivariate time series based on candlestick charts, recording six dimensions of data within fixed time intervals: open, high, low, close prices, plus volume and turnover. These sequences form a highly compact, information-dense “language” that market participants use to interpret price fluctuations, volatility states, liquidity changes, and shifts in collective sentiment.

However, for multivariate time-series models—whether classical econometric models like ARIMA, machine learning models like LSTM, or existing general time-series foundation models like TimeGPT—their approach to data is essentially numerical computation. In these models’ “perspective,” the six dimensions of K-line data, though input simultaneously, are merely treated as a set of multidimensional floating-point vectors. The models lack a mechanism to recognize these six dimensions as a ‘semantic whole (K-line pattern).’ This underlying logic of decomposing K-lines into pure numbers often prevents models from capturing structured features with strong financial meanings, such as “long lower shadows.”

To address these limitations, the Li Jian team at Tsinghua University launched Kronos in August this year. Kronos is a unified, scalable pre-training framework built on a decoder-only Transformer architecture, specifically designed for financial K-line data. For the first time, the Kronos universe treats K-line data as a logical “market language.” Here, K-lines and their combinations are no longer isolated numerical sequences; they become carriers expressing market states, capital intent, and trend directions, possessing semantics, syntax, and contextual dependencies.

Before diving into Kronos’s workflow, let’s discuss its naming. The name reflects the research team’s ingenuity and ambition. In Greek mythology, Kronos is the god of time, symbolizing the power to impose order on chaos and control the flow of time. By naming the model “Kronos,” the team aims to reveal its core objective: to become a “time master” capable of understanding the chaotic fluctuations of financial time series and extracting patterns from them.

Deep Dive: How Kronos Imparts “Semantics” to K-Lines

Kronos’s architecture is clear, consisting of two core components: the “Tokenizer” and the “Autoregressive Model,” depicted in the left and right panels of the figure below. Let’s first focus on the first part—the tokenizer. The goal of this stage is to transform continuous K-line data into machine-understandable “financial semantic units (Tokens).”

In NLP, tokens are the smallest units of language. In Kronos’s system, the smallest unit of financial market language is a “semantic unit,” composed of one or more K-lines. However, raw K-lines contain six dimensions of continuous data (open, high, low, close, volume, turnover) and cannot directly serve as semantic “market language units.” Therefore, the research team designed a specialized tokenizer for K-lines.

It first feeds the raw K-line sequence (the red and green K-lines at the top of the diagram) into the “Tokenizer Encoder.” Using Binary Spherical Quantization (BSQ) technology, it quantizes the six-dimensional continuous data into layered, discrete Tokens. To capture both the market’s “macro trends” and “micro details,” Kronos employs a hierarchical Token design, splitting each Token into coarse-grained sub-Tokens and fine-grained sub-Tokens.

Coarse-grained sub-Tokens provide a low-fidelity reconstruction of the original continuous K-line data, responsible for capturing macro features like “price trends” and “volume magnitude.” Fine-grained sub-Tokens represent the residuals or detailed modifications of the coarse-grained representation, responsible for capturing micro market structures, precise amplitude values, and high-frequency volatility details. Simply put, coarse grains handle the big picture, while fine grains handle the details.

The quantized Tokens then pass through the “Tokenizer Decoder” to reconstruct the K-line sequence (Reconstruction at the bottom of the diagram), ensuring that the Tokens are both discretized and retain the core information of the original K-lines.

Here lies a technically profound highlight. Financial data is continuous, theoretically allowing infinite values and forming an “infinite state space.” Traditional discretization methods, such as equal-interval binning or fixed ranges, forcibly split cases that are “semantically similar but numerically slightly different” (e.g., 100.1 vs. 100.2). Furthermore, the infinite state space leads to parameter overload, resulting in high computational complexity and失效 (failure) of generalization capabilities.

Therefore, Kronos introduces Binary Spherical Quantization (BSQ) technology. It projects high-dimensional vectors onto a hypersphere to find similar projections. This means that even if price values fluctuate slightly, if their patterns and directions are similar, they are encoded as the same Token. This process is akin to treating “happy” and “joyful” as synonyms in natural language. This is the source of Kronos’s noise resistance. Through this sophisticated tokenizer, Kronos has successfully compiled an exclusive “K-Line Dictionary” for financial markets.

With exclusive “financial words” (Tokens), the next step is to teach the model to “read”—Autoregressive Pre-training. The goal of this stage is to enable the model to learn the “sequential logic of market language Tokens.” Kronos adopts an autoregressive pre-training objective consistent with GPT. Thanks to its Decoder-only architecture, Kronos inherently possesses generative capabilities. It predicts the next most likely “financial word” based on a sequence of past “financial vocabulary.”

During Kronos’s pre-training or prediction, the autoregressive model strictly follows sequential dependencies: Step 1: Predict the coarse-grained sub-Token ($k_c$) of the next Token based on historical Token information (e.g., the “3” output in the Header);

Step 2: Further predict the corresponding fine-grained sub-Token ($k_f$) (e.g., the “4” output in the Header) by combining the predicted coarse-grained information via “Cross Attention.”

To train this “brain,” the research team fed it a massive corpus containing over 12 billion K-line records from 45 global exchanges, including XSHG (Shanghai Stock Exchange), XNAS (NASDAQ), XJPX (Tokyo Stock Exchange), Crypto, and Forex. In this process, Kronos learned the “universal grammar” of financial markets, mastering common rise-and-fall patterns across assets and regions.

The empirical data in the paper convincingly validates its zero-shot transfer capability. When Kronos, pre-trained on multi-market corpora from 45 global exchanges (including US stocks/XNAS), is directly applied to predict unseen China A-shares (XSHG) markets, its prediction accuracy (RankIC) decays by only 5%-10%. In contrast, traditional time-series models suffer a performance decay of 35%-45% in the same cross-market zero-shot scenario. The essence of this vast difference is that traditional models rely on “numerical fitting,” learning only the numerical distribution patterns of specific markets. Kronos, however, masters cross-market universal financial semantic logic. These semantic patterns, determined by market participants’ behavioral logic, are consistent across different markets, thereby offering stronger generalization capabilities.

From Backtesting to “Parallel Universes”: Kronos’s Practical Imagination

As mentioned earlier, Kronos possesses synthetic data capabilities, which we can leverage.

The biggest pain point of traditional backtesting is that history occurs only once. If a strategy made money in the past, was it due to robust logic or merely luck, coinciding with a favorable historical period? We often cannot know. Kronos now provides us with a laboratory condition. You can ask it to generate 1,000 “unoccurring but theoretically possible” subsequent price paths based on data from January 2020. If your strategy makes huge profits in real history but loses miserably in 600 of Kronos’s 1,000 “parallel universes,” it indicates your strategy was merely “lucky” and caught the right market segment. Conversely, if it performs robustly in 90% of the parallel universes, we can confidently say: this strategy’s logic is genuinely robust. This represents a qualitative shift from “backtesting” to “stress testing.” It can address the “survivorship bias” problem in quantitative strategies.

Furthermore, Kronos’s open-source nature may offer individual developers and small-to-medium institutions an efficient path of ‘small-sample real data + large-scale synthetic data.’ We only need to purchase recent real high-frequency data as a ‘seed’ (Prompt) to ensure logical validity, allowing Kronos to generate infinite variant data at low cost based on this seed. This means developers do not need to purchase expensive, ultra-long historical data. By using a small amount of real data as a starting point, they can build a massive training set covering various extreme market conditions and volatility regimes through Kronos. This significantly lowers the data barrier for quantitative research, enabling ‘data-poor’ individual developers to train deep learning strategies with strong generalization capabilities.

Controversy and Reflection: Is Pre-trained Models the Endgame for Quant?

Criticism primarily centers on whether Kronos has live-trading value. Some are shocked that Kronos’s out-of-sample testing used only a short period in 2024. They argue that a true quantitative strategy must span bull and bear markets, requiring 5 to 10 years of backtesting to verify robustness. Data from 2024 alone cannot prove a strategy’s survival capability in extreme environments like “circuit breakers,” “stock market crashes,” or “massive liquidity injections.” Others contend that the paper discusses only MSE (Mean Squared Error) and MAE (Mean Absolute Error) without mentioning backtest return curves, suggesting the paper was published merely for academic credit and that live trading would inevitably result in losses.

From a practical standpoint, their skepticism is justified. If we do not acknowledge this, we are blindly praising the model.

However, from an AI research perspective, the evaluation metric in large language model research is typically token prediction accuracy. For high-frequency/minute-level K-lines, one year of data contains millions of tokens, which is statistically sufficient to verify “whether the model has converged.” Moreover, Kronos is essentially a generative model; its task is to ‘reconstruct the market,’ not to ‘beat the market.’ Low MAE and MSE indicate that the model understands market volatility patterns well and knows where prices are likely to oscillate next. This does not, however, mean it can precisely capture Alpha.

It is akin to allowing GPT to write perfect Python code based on requirements. If you ask GPT-4 to predict which stock will rise, it is essentially guessing blindly. Therefore, using a pre-trained model directly for trading will inevitably lead to losses. This does not mean the model lacks value; it only means the approach is incorrect.

In summary, Kronos reconstructs K-line sequences into semantically rich sentences, turning each K-line into a ‘word’ (Token) carrying information. It brings a new perspective to time-series research.

This Issue’s Question

For this issue’s cover image, we used a photo from a certain university. Guess which university it is.