Trading with the Devil: Risk and Return in Foundation Model Strategies

TL;DR

This paper extends the Capital Asset Pricing Model (CAPM) to trading strategies built on foundation models by decomposing risk into a shared epistemic component from the pretrained backbone and an idiosyncratic aleatory component from fine-tuning. It introduces the Pretrained Market Line (PML) and the Foundation Sharpe Ratio as tools to quantify the platform-wide risk-return frontier, using Monte Carlo dropout to separate priced from unpriced risk. Empirical validation on US equities and cryptocurrencies shows that the shared model risk aligns with systemic risk in CAPM, while alpha potential decays as the market incorporates these signals. The framework provides a transparent lens for evaluating foundation-model trading and highlights practical considerations for latency, cross-model risk, and the safe scaling of such models in competitive financial markets.

Abstract

Foundation models - already transformative in domains such as natural language processing - are now starting to emerge for time-series tasks in finance. While these pretrained architectures promise versatile predictive signals, little is known about how they shape the risk profiles of the trading strategies built atop them, leaving practitioners reluctant to commit serious capital. In this paper, we propose an extension to the Capital Asset Pricing Model (CAPM) that disentangles the systematic risk introduced by a shared foundation model - potentially capable of generating alpha if the underlying model is genuinely predictive - from the idiosyncratic risk attributable to custom fine-tuning, which typically accrues no systematic premium. To enable a practical estimation of these separate risks, we align this decomposition with the concepts of uncertainty disentanglement, casting systematic risk as epistemic uncertainty (rooted in the pretrained model) and idiosyncratic risk as aleatory uncertainty (introduced during custom adaptations). Under the Aleatory Collapse Assumption, we illustrate how Monte Carlo dropout - among other methods in the uncertainty-quantization toolkit - can directly measure the epistemic risk, thereby mapping trading strategies to a more transparent risk-return plane. Our experiments show that isolating these distinct risk factors yields deeper insights into the performance limits of foundation-model-based strategies, their model degradation over time, and potential avenues for targeted refinements. Taken together, our results highlight both the promise and the pitfalls of deploying large pretrained models in competitive financial markets.

Figures (4)

• Figure 1: A High‐Level View: From the Classical CAPM to the “Foundation‐Model” CAPM. On the left, we show the familiar Capital Market Line (CML) linking the risk‐free asset $R_f$ to a tangential portfolio $M$, which represents the fully diversified market portfolio in CAPM theory. On the right, we illustrate our foundation‐model extension, where the tangential portfolio $M$ is replaced by the tangential strategy $\theta_T$ that fully exploits the pretrained network. Custom fine‐tuned strategies $\theta_s$ may deviate from the Pretrained Market Line (PML) if they bear additional, unpriced (idiosyncratic) variance. We posit a hypothetical optimal strategy $\theta_{\mathrm{opt}}$, which shares the same expected return but carries only epistemic risk. The distance in standard deviation between $\theta_s$ and $\theta_{\mathrm{opt}}$ captures the extra variance that could be avoided by more efficient adaptation of the foundation model. This figure foreshadows our framework for adapting CAPM principles to foundation‐model trading, forming the conceptual basis for Sections 3.1 and 3.2.
• Figure 2: TimesFMv2's model surprise–market return correlation at two prediction resolutions for cryptos. Top: Correlation coefficients for an ultrafast (100ms) prediction horizon, capturing rapid market reactions. Bottom: Corresponding coefficients for a 1 day horizon, illustrating how the model surprise metric behaves at extended timescales.
• Figure 3: Daily Risk–Return outcomes (in bps) for all evaluated strategies. (a) Grey markers represent non‐pretrained baseline strategies, and colored markers correspond to foundation‐model–based approaches. (b) A more granular view of Chronos at five parameter scales (8M to 710M), illustrating how model size influences risk–return trade‐offs.
• Figure 4: Rolling‐window Sharpe ratios for a TimesFM‐v1‐based strategy, re‐fine‐tuned weekly, from early 2023 to early 2025. The maroon curve ( $\mathrm{SR}{\theta_s}$ ) represents the observed performance, while the gold curve ( $\mathrm{SR}_{\theta}$ ) reflects the theoretical upper bound from PML estimation. Vertical dashed lines denote the public release dates of TimesFM‐v1 and TimesFM‐v2.

Theorems & Definitions (2)

• definition 1: Backbone family
• definition 2: Tangential backbone strategy