📊 statistics

Thick as THieFs: Temporal coherent forecast combination for day-ahead electricity prices

This paper proposes a temporal coherent forecast combination method that simultaneously resolves incoherence across different temporal granularities and leverages the strengths of multiple competing models to produce minimum-variance, unbiased day-ahead electricity price forecasts that significantly outperform existing approaches in German and Spanish markets.

Original authors: Daniele Girolimetto

Published 2026-07-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daniele Girolimetto

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather for a whole week. You have a team of experts: one looks at satellite clouds, another studies wind patterns, and a third checks historical temperature data. If you ask each of them for a prediction, they might give you different answers. But here's the tricky part: if you ask them for the temperature of every single hour, and then ask for the average temperature for the whole day, their answers might not add up. Maybe the sum of their hourly guesses doesn't match their daily guess. That's a problem. In the world of energy, specifically electricity, this isn't just a math puzzle; it's a money-maker (or a money-loser). Electricity prices change every hour, but traders also buy "blocks" of time (like 4-hour chunks) or "baseload" contracts (the average price for the whole day). If your predictions for the hours don't match your prediction for the day, you might make a bad trade. This paper tackles how to fix these mismatched predictions and how to combine the best ideas from different experts into one super-accurate forecast.

The paper, titled "Thick as THieFs," introduces a clever method to solve two headaches at once. First, it fixes the "incoherence" problem where hourly predictions don't match daily averages. Second, it stops us from having to pick just one "best" model. Instead of betting on a single crystal ball, the authors combine the forecasts from four very different "experts" (a simple statistical model, a neural network, a tree-based learner, and a massive AI foundation model) into a single, unified prediction. They call this "Temporal Coherent Forecast Combination." Think of it like a choir. If you have four singers, each with a different voice, you don't just pick the loudest one. Instead, you blend their voices together, but you make sure they all hit the right notes so the song sounds perfect from start to finish. The authors tested this on electricity markets in Germany and Spain and found that this "super-choir" approach was significantly more accurate than any single singer or even any single singer who had been tuned up to match the group.

Here is how the magic works. The authors realized that electricity prices exist at different "levels" of detail: the 24 individual hours of a day, blocks of 2, 3, 4, 6, 8, or 12 hours, and finally the single average for the whole day. These levels are mathematically linked; the daily average must be the sum of the hours. Traditional methods often predict each level separately, leading to the mismatch problem. The paper proposes a two-step fix wrapped into one. First, they take the raw, messy predictions from four different models. Second, they use a mathematical "reconciliation" process to force these predictions to obey the rules of the hierarchy (so the hours add up to the day) while simultaneously blending the four experts' opinions to minimize errors.

The results are quite impressive. When they tested this method on real data from 2021 to 2024, the combined forecast beat the best individual expert in almost every scenario. For the hourly prices, the new method reduced the error (measured by Mean Absolute Error) by between 4% and 24% compared to the original, un-blended forecasts. The gains were even bigger for the longer time blocks, like the 12-hour or 24-hour averages, which are crucial for traders. The authors also discovered that the most important part of their recipe wasn't the fancy math used to smooth out the data, but rather how they handled the relationship between the different experts. Preserving the fact that experts might make similar mistakes at the same time was key to success.

Interestingly, the paper argues against the idea that you need to pick the single "smartest" model. Even the most advanced AI model in their study (called MITRA) performed better when its predictions were combined with the others. The paper also rules out the idea that you should first average the experts and then fix the math. They found that doing both steps at the same time—combining and reconciling in a single "simultaneous step"—is the superior strategy. While the method relies on complex statistics to estimate how the errors of different models relate to one another, the core finding is simple: when you have a hierarchy of predictions and a team of experts, blending them all together while respecting the rules of the game yields a forecast that is harder to beat. This isn't just a theoretical win; it offers a practical, robust way to predict electricity prices more accurately, which helps keep the lights on and the markets stable.

Technical Summary: Thick as THieFs: Temporal coherent forecast combination for day-ahead electricity prices

Problem Statement
Day-ahead electricity prices are forecast using a variety of competing models across multiple temporal granularities, ranging from hourly prices to block and baseload products. This setup presents two distinct challenges:

Incoherence: Forecasts produced at different temporal granularities often fail to satisfy aggregation constraints (e.g., the average of predicted hourly prices does not match the predicted daily baseload price).
Model Uncertainty: No single forecasting model is consistently the most accurate across all markets, time periods, and temporal levels.

Existing literature addresses these issues separately. Temporal reconciliation (e.g., Athanasopoulos et al., 2017) ensures coherence for a single model across temporal levels but does not combine competing models. Conversely, forecast combination (e.g., Bates and Granger, 1969) aggregates multiple models but typically ignores the structural constraints imposed by temporal hierarchies.

Methodology
The paper proposes a unified framework that simultaneously addresses incoherence and model uncertainty by deriving a temporal coherent forecast combination.

Framework: The approach builds upon the coherent forecast combination framework for cross-sectionally constrained time series (Girolimetto and Di Fonzo, 2024) and extends it to temporal hierarchies. It considers $Q \geq 1$ experts (forecasting models), each producing base forecasts for all levels of a temporal hierarchy.
Optimization: The method seeks a linear, unbiased combination of the stacked base forecasts that satisfies the temporal aggregation constraints and minimizes the error variance.
Closed-Form Solution: The optimal combination is derived in two equivalent forms:
1. Structural Form: $\hat{z}_{c} = S(S_Q^\top W^{-1} S_Q)^{-1} S_Q^\top W^{-1} \hat{z}$ , where $S$ is the structural matrix mapping high-frequency data to the hierarchy, $S_Q$ replicates this for $Q$ experts, and $W$ is the covariance matrix of the stacked base forecast errors.
2. Zero-Constrained Form: A linearly constrained least-squares adjustment.
- This solution generalizes the single-expert temporal reconciliation of Athanasopoulos et al. (2017) as the special case where $Q=1$ .
Covariance Estimation: Since the optimal solution depends on the high-dimensional error covariance matrix $W$ $W$ , which is unknown, the paper evaluates various estimation strategies. These strategies combine:
- Correlation Structures: Four structures ranging from "Full" (no constraints) to "Diagonal" (no correlations), including "By-expert" (preserving intra-expert temporal dependence) and "By-node" (preserving inter-expert dependence at specific time points).
- Shrinkage Techniques: Linear shrinkage (Ledoit and Wolf, 2004) and three nonlinear shrinkage methods (Ledoit and Wolf, 2022a) to regularize the estimation of the covariance matrix, ensuring it is positive-definite and well-conditioned.

Experimental Setup
The methodology was evaluated using day-ahead electricity price data from the German (EPEX-DE) and Spanish (OMIE) markets over the 2021–2024 period.

Data: The study utilized base forecasts from Lipiecki et al. (2026), which cover four distinct model classes: ARX (linear autoregressive), NARX (nonlinear neural networks), XGB (gradient-boosted trees), and MITRA (pretrained foundation model).
Hierarchy: The temporal hierarchy consisted of 8 levels ( $K = \{24, 12, 8, 6, 4, 3, 2, 1\}$ ), representing hourly prices up to the daily average.
Comparison: The proposed coherent combination was compared against:
1. The original base forecasts.
2. Single-expert temporal reconciliation (reconciling each model independently).
3. A sequential strategy (averaging experts first, then reconciling).

Key Results

Superior Accuracy: The temporal coherent combination significantly outperformed both the base forecasts and the best single-expert reconciled forecast at nearly every temporal level in both markets.
Error Reduction: Relative to base forecasts, the combination reduced hourly Mean Absolute Error (MAE) by 4% to 24%, depending on the expert. Gains were generally larger at coarser temporal aggregations (e.g., block and baseload products).
Statistical Significance: Diebold–Mariano tests confirmed that the combined forecasts were significantly more accurate than base and reconciled forecasts in the vast majority of cases.
Robustness:
- The choice of correlation structure was found to be the critical modeling decision. Preserving each expert's error dependence across temporal levels (the "by-expert" structure) was substantially more important than the specific linear or nonlinear shrinkage method used.
- The gains were not driven by a single expert; the combination improved upon the best individual expert in almost all scenarios.
- The simultaneous approach (combining and reconciling in one step) outperformed the sequential strategy (averaging then reconciling).

Significance and Claims
The paper claims that temporal coherent combination offers a practical and statistically rigorous solution to the dual problems of incoherence and model selection in electricity price forecasting. By treating the temporal hierarchy and the ensemble of experts jointly, the method fully exploits available information along two dimensions: temporal aggregation levels and forecasting models.

The authors emphasize that this approach provides a "single forecast" that satisfies aggregation constraints and minimizes error variance among linear unbiased combinations. The results suggest that for day-ahead electricity markets, where decisions are made at various temporal granularities, a unified coherent combination is superior to selecting a single "best" model or reconciling models individually. The study validates the extension of the "ThieF" (Temporal Hierarchical Forecasting) concept from single-model reconciliation to multi-model combination, demonstrating that foundation models and traditional statistical models can be effectively integrated to improve forecast accuracy across the entire temporal spectrum.

More like this