Falsifiable Compositional Reasoning in Time Series Forecasting
Organizations: Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China.
Abstract
Time series foundation models as well as the language-guided analytical agents achieve strong accuracy on standard chronological benchmarks. However, models that rank highest on historical validation splits often degrade across non-stationary regime transitions. We hypothesise that this gap arises from using historical backtesting as an open-ended search and model-selection objective instead of using it to test pre-specified compositional assumptions. Real-world time series are generated by recurring latent primitives that co-move within finite historical windows. Under strong historical co-movement, chronological holdout error cannot separate predictors that learn invariant couplings among primitives from predictors that fit transient historical correlations. Searching over large pools of architectures, hyperparameters, or chain-of-thought rules further biases selection toward historical shortcuts. Through theoretical analysis and diagnostic experiments on a controlled multi-primitive benchmark and German day-ahead electricity prices, we show that expanding validation search fails to improve out-of-regime accuracy, and that decomposing primitives without an ex-ante falsification test still breaks down when residual correlations shift. Hence, we propose formulating time series forecasting and reasoning around falsifiable compositional hypotheses that combine neural primitive encoders, explicit coupling constraints, and training-calibrated falsification tests.
Figures & tables
| Method | ID MSE | Shift MSE | Reject % |
|---|---|---|---|
| Controlled Multi-Primitive Benchmark | |||
| End-to-End PatchTST | |||
| Unguided Rule/CoT | |||
| Comp. (w/o test) | |||
| Comp. (w/ test) | |||
| German Electricity Transition | |||
| Pre-Committed Element | Methodological Role | Example A: Day-Ahead Grid Pricing ( EPF-DE ) | Example B: Intraday Order-Flow Returns |
|---|---|---|---|
| 1. Grounded Primitives | Isolate reusable temporal drivers from entangled target lags. | : 72h de-effected baseload level, : grid load forecast, : combined wind and solar supply. | : dealer inventory imbalance, : macro-announcement volatility state, : cross-venue spread. |
| 2. Persistence Mechanism | Specify the physical or institutional mechanism sustaining the law. | Power-balance clearing along thermal merit-order dispatch curves ( Weron, 2014 ) . | Dealer inventory risk control requiring intraday liquidity compensation ( Kyle, 1985 ; Cont et al., 2014 ) . |
| 3. Directional Coupling | Constrain signs, symmetries, or conservation relations across primitives. | Monotone positive response of to net residual load , steepening near reserve limits. | Negative return drift proportional to dealer inventory , scaled inversely by depth . |
| 4. Regime Boundary Conditions | Define observable states where switches functional form or clips. | Curtailment and negative-price regime when renewable supply exceeds must-run thermal capacity. | Exchange price-limit halts, closing-auction cutoffs, and overnight margin liquidation windows. |
| 5. Ex-Ante Falsification Test | Specify a training-calibrated support test that shrinks invalid residual fits. | 4D Mahalanobis test on , , net load , and attenuating diurnal lag residuals. | Rolling likelihood-ratio test on inventory-return sign consistency and spread jumps attenuating alpha residuals. |
Time Series Forecasting
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
See all 947 papers