Forecast Collapse: When Foundation Models Go Flat

The Problem

Financial forecasting is the original time-series killer app. Hedge funds, trading desks, and portfolio managers all want the same thing: a model that can rank tomorrow's expected returns across a universe of stocks. If your model tells you stock A will outperform stock B, that's actionable. If it tells you both will return 0.02%, you have nothing.

A new paper from researchers at several institutions (Wan, Ma, Zhu, Liu, Wang, Wen, Liu) documents a phenomenon they call forecast collapse: when foundation models and deep learning architectures forecast hourly returns for 1,000 US equities, the predictions go nearly flat. Every stock gets roughly the same number. Cross-sectional rank correlation drops toward zero. The model becomes useless for the one thing a financial forecast must do — tell you which stocks to buy.

And then the kicker: forecast the same equities' trading volume under identical settings, and the collapse disappears. The models work fine. The problem is specific to returns, and it's pervasive.

The Method

The authors run a systematic investigation across an unusually broad array of models: time-series foundation models (TSFMs like TimesFM, PatchTST, Informer), twelve deep-learning forecasting architectures total, and 97 public benchmark configurations. The primary testbed is Finance1K, a dataset of hourly returns and volumes for 1,000 US equities.

They evaluate along two axes that normally live in separate worlds:

• Per-series calibration — does each stock's forecast have low mean squared error relative to its own history?
• Cross-sectional ranking — do the forecasts correctly order stocks by next-period return?

The discovery is that these axes are in direct tension. A model can be well-calibrated per-series and simultaneously useless for ranking. The paper traces the collapse to two distinct mechanisms.

graph TB
    subgraph "Forecast Collapse Mechanisms"
        M1[Low Signal-to-Noise in Returns Data]
        M2[Per-Series MSE Objective]
        M3[Cross-Series Structure Left Unidentified]
    end
    subgraph "Consequence"
        C1[Flat Predictions]
        C2[Poor Stock Ranking]
    end
    subgraph "Solution: CalibRank"
        R1[Balance Calibration and Ranking]
        R2[Nearly 3x Cross-Sectional Correlation]
    end
    M1 --> C1
    M2 --> C1
    M3 --> C2
    C1 & C2 --> R1
    R1 --> R2
    style C1 fill:#7f1d1d,stroke:#ef4444
    style C2 fill:#7f1d1d,stroke:#ef4444
    style R1 fill:#1e3a5f,stroke:#3b82f6
    style R2 fill:#1e3a5f,stroke:#3b82f6

The Numbers

The central empirical result is stark. Under standard per-series MSE objectives, cross-sectional rank correlation across 1,000 equities hovers near zero for returns forecasts. The predicted values cluster so tightly that the variance across stocks is a fraction of the true return variance.

• Calibration-ranking tradeoff: Optimizing squared error drives forecasts toward the conditional mean per series — which, for low-predictability returns, is nearly constant. Optimizing rank correlation directly inflates forecast amplitude by "more than an order of magnitude" and overstates predictability.
• Volume vs. returns: Trading volume, which has higher signal-to-noise ratio, does not exhibit forecast collapse under identical model configurations. The phenomenon is tied to target predictability, not a model architecture flaw.
• CalibRank: A proposed loss that combines per-series MSE with a cross-sectional rank correlation term. On Finance1K, it "nearly triples cross-sectional correlation while keeping amplitude close to the target" and improves correlation across all tested models.

The paper tests this across 97 benchmark configurations and finds the collapse is not a quirk of one architecture or dataset. It's structural.

Limitations

Three caveats are worth highlighting.

First, Finance1K is hourly equity returns. The authors don't test whether forecast collapse generalizes to other asset classes (crypto, FX, fixed income) or frequencies (daily, weekly, tick-level). Returns at other horizons may have very different signal structure.

Second, CalibRank is evaluated on historical data, not trading. Improving cross-sectional correlation is a proxy for better ranking, but the paper doesn't simulate a trading strategy to measure whether CalibRank's forecasts translate into actual P&L. Rank correlation and trading returns are related but not identical.

Third, the collapse is documented for point forecasts. Probabilistic forecasting — predicting a distribution of returns rather than a single number — may behave differently. The paper doesn't explore whether forecast collapse extends to quantiles or density predictions, where the flatness problem could be less severe or completely different.

Why This Matters

This paper reveals a blind spot in how the time-series community evaluates models. Per-series metrics like MSE dominate the literature, but they can hide a complete failure in the cross-series structure that downstream decisions need. A model that scores well on held-out MSE but produces flat predictions across stocks is worse than useless — it's misleading.

For anyone building with time-series foundation models, the implications are immediate:

• Always check cross-sectional rank correlation even if your primary metric is per-series MSE. The two can diverge dramatically.
• Your loss function encodes your priorities. CalibRank shows that a simple modification to the objective can recover ranking ability without sacrificing calibration. Not all problems need a new architecture — sometimes they need a new loss.
• Low-predictability targets expose hidden failures. If your time-series model works on volume but collapses on returns, it's not the model's fault — it's your evaluation that's incomplete.

This is a quiet but important paper. It doesn't propose a new state-of-the-art architecture. It reveals that a standard evaluation practice is hiding a failure mode — and offers a fix that costs nothing but a term in the loss function. That's the kind of research that actually changes how people build.