A stronger forecast is only the beginning.
Foundation models, changing market regimes, and the distance between a prediction and a tradable system.
| Section | Sources cited | Equations | Code listings |
|---|---|---|---|
| 1. Why point forecast accuracy diverges from realized portfolio utility | 1, 2 | 2 | 0 |
| 2. Validate the full forecast-to-allocation chain on synthetic data | 1, 3 | 0 | 1 |
01 /Why point forecast accuracy diverges from realized portfolio utility
Time-series foundation models bring pretrained representations to financial forecasting. A June 2026 benchmark compares models including Chronos, TimesFM, and Moirai with neural baselines and random-walk forecasts on five liquid US equities [1]. Its reported gains over the random walk are small and sparse; forecast rankings alone do not establish a profitable strategy after trading costs.
Prediction quality, portfolio behavior, and execution are distinct mathematical operators. A model that reduces mean squared error on smooth price levels can still generate excessive sign flips in expected returns, inducing a turnover tax that dominates gross alpha. A signal must survive its transformation into bounded positions and clear the bid-ask and impact hurdle.
Rather than feeding raw point forecasts into an unconstrained allocator, we extract calibrated predictive quantiles and define an inter-quantile dispersion measure . Combined with causal regime filtering [2], predictive uncertainty gates position sizing directly.
Evaluated out-of-sample across τ ∈ {0.16, 0.50, 0.84} to audit both central tendency and dispersion calibration.
This illustrative shrinkage rule sets the target weight to zero when |q̂_{0.50,t}| ≤ c_t. An existing position may still need liquidation. The threshold is a modeling choice, separate from the one-way turnover cost used to evaluate trades.
02 /Validate the full forecast-to-allocation chain on synthetic data
A practical research program compares a foundation model with local statistical baselines across identical forward folds [1, 3]. It tests the complete conversion from quantile distribution to net portfolio return rather than stopping at RMSE or MAE tables.
We design around that complete chain: data hygiene, pretraining cutoff audits, quantile calibration, uncertainty-scaled sizing, risk budgets, and execution monitoring. In the seeded synthetic scenario below, shrinking small forecasts improves the net sum of returns after turnover costs. This single path does not establish that buffering improves every strategy or market regime.
- Audit foundation-model pretraining cutoffs to prevent historical benchmark memorization from masquerading as forward skill.
- Score predictive quantiles with pinball loss and coverage ratios before computing any portfolio weights.
- Couple median forecasts with an explicit no-trade zone scaled to instrument-level spread and impact estimates.
# Illustrative only; compares unbuffered vs friction-buffered quantile sizing.
import numpy as np
rng = np.random.default_rng(42)
T = 400
true_drift = 0.0008 * np.sin(np.linspace(0, 8 * np.pi, T))
realized = true_drift + rng.normal(0.0, 0.01, size=T)
q50 = true_drift + rng.normal(0.0, 0.0015, size=T)
dispersion = np.full(T, 0.02)
signal_threshold = 0.0010 # Decimal return units: 10 basis points.
one_way_cost = 0.0010
raw_w = np.clip(q50 / (2.0 * dispersion**2), -1.0, 1.0)
buffered_w = np.clip(np.sign(q50) * np.maximum(np.abs(q50) - signal_threshold, 0.0) / (2.0 * dispersion**2), -1.0, 1.0)
net_raw = float(np.sum(raw_w * realized - one_way_cost * np.abs(np.diff(np.r_[0.0, raw_w]))))
net_buffered = float(np.sum(buffered_w * realized - one_way_cost * np.abs(np.diff(np.r_[0.0, buffered_w]))))
assert net_buffered > net_raw
print(f"net_raw={net_raw:.4f}, net_buffered={net_buffered:.4f}")Research is complete only when the assumptions, experiment, and limitations can be reconstructed.
This perspective is research commentary. It does not constitute investment advice or evidence of live strategy performance. Experimental methods require project-specific testing. Equations and code illustrate research formulations, not measured live returns.