What concrete gains can you expect in forecasting and risk analysis?
The most immediate impact will be a step-change in forecast accuracy, especially in zero-shot settings where a model is applied to a new market or asset without retraining. The Kronos model, pre-trained on over 12 billion K-line records from 45 global exchanges, improved price series forecasting RankIC by 93% over the leading time-series foundation model and 87% over the best non-pre-trained baseline [1]. In plain terms, that means a model that has never seen a particular stock can still predict its price movements far more accurately than previous best-in-class methods.
Risk measurement also stands to benefit. The same Kronos model reduced volatility forecasting error by 9% and improved the realism of synthetic K-line data by 22% [1]. Better volatility forecasts directly improve Value-at-Risk (VaR) estimates, which banks and funds use to gauge potential losses. A separate 2025 study combined a time-series foundation model with historical simulation to forecast crude oil price risk, finding improved reliability [2]. This suggests that foundation models can capture the complex, nonlinear dynamics that traditional risk models often miss.
How do these models achieve such gains?
The secret is scale and transfer learning. Instead of building a separate model for each dataset, foundation models are pre-trained on enormous, diverse financial corpora. For instance, Kronos uses a specialized tokenizer that converts continuous price and volume data into discrete tokens, similar to how language models process words, and then learns to predict the next token [1]. This allows the model to capture both price dynamics and trading activity patterns, which are crucial for financial forecasting.
Other architectures, such as the LSTM-Transformer hybrid, combine the strengths of different neural networks to handle the noisy, nonlinear nature of financial data [4]. The key is that these models learn generalizable patterns from vast amounts of data, which they can then apply to new tasks with minimal fine-tuning. This is a fundamental shift from the traditional one-dataset-per-model approach, as highlighted in a 2025 survey of time-series foundation models in finance [3].
What are the catches and limitations?
Despite the promise, there are significant hurdles. A 2025 survey warns that claims of 'better forecasts' must be inseparable from tail-risk adequacy—meaning a model might be accurate on average but fail to predict extreme market moves [3]. The survey recommends pairing standard probabilistic scoring with industry-standard VaR/ES backtests, such as Kupiec's and Christoffersen's tests, to ensure models are reliable in crisis scenarios [3]. This is a critical reminder that accuracy in normal times does not guarantee safety in turbulent ones.
Another catch is the risk of data-snooping, where a model appears to perform well simply because it was tested on the same data it was trained on. The survey advocates for rigorous frameworks like White's Reality Check and Hansen's SPA test to guard against this [3]. Additionally, while foundation models are powerful, they are not always superior to simpler hybrid approaches. A 2025 study found that a hybrid LSTM-ARIMA model achieved 92.7% risk prediction accuracy, outperforming either model alone [5]. This suggests that combining AI with traditional statistical methods can still be highly effective, especially when data is limited.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2025 to 2026, 5 from 2024 or later, 1 in Q1 journals, collectively cited 53 times — selected as the most relevant from 8 studies that passed quality screening, drawn from 55 papers retrieved from a database of over 500 million.
Sources used in this answer
Kronos: A Foundation Model for the Language of Financial Markets
Kronos, a foundation model pre-trained on over 12 billion K-line records from 45 exchanges, improved price forecasting RankIC by 93% over the leading TSFM and 87% over the best non-pre-trained baseline, while also reducing volatility forecast MAE by 9% and improving generative fidelity by 22%.
Crude oil risk forecasting using time series foundation model
A proposed Value Risk forecasting model using a time-series foundation model combined with historical simulation showed improved reliability in forecasting crude oil price risk.
Time-Series Foundation Models in Finance: Pretraining Corpora, Architectures, Financial Benchmarks, and Risk-Aware Evaluation
A 2025 survey of time-series foundation models in finance highlights the need for discipline-specific evaluation protocols that include probabilistic scoring (CRPS, pinball loss) and tail-risk backtests (Kupiec, Christoffersen, Acerbi-Székely) to ensure models are reliable for risk management.
LSTM–Transformer-Based Robust Hybrid Deep Learning Model for Financial Time Series Forecasting
An LSTM-Transformer-MLP hybrid model demonstrated exceptional forecasting performance and robustness across multiple financial datasets, including Bitcoin and stock indices, compared to benchmark and state-of-the-art models.
Risk Prediction in Financial Markets Using Hybrid AI and Time Series Forecasting Models
A hybrid LSTM-ARIMA model improved risk prediction accuracy to 92.7%, outperforming either model alone, as measured by RMSE and MSE on historical market data.