TriForecaster: Triadic Forecasting Models

• TriForecaster is a design pattern that organizes forecasting systems into three interacting components, addressing challenges like heterogeneity and long-range dependencies.
• It leverages techniques such as triangular attention, tri-modal fusion, and three-stage post-processing to enhance efficiency, calibration, and model specialization.
• Empirical evaluations, including a 22.4% MSE reduction in electric load forecasting, demonstrate its practical gains across diverse time-series applications.

TriForecaster denotes a class of forecasting systems organized around three coordinated components, modalities, stages, or specialization axes. In the provided literature, the label spans long-sequence multivariate time-series models built from triangular patch attention and variable-specific parameterization (Cirstea et al., 2022), tri-modal architectures that jointly encode time, frequency, and prompt or vision spaces (Chowdhury et al., 6 Aug 2025, Ao, 17 Apr 2026), sample-aware models with Temporal-Spatial-Sample attention (Chen et al., 5 Jun 2026), tri-stage probabilistic post-processing pipelines for weather forecasts (Kirkwood et al., 2020), tri-component decomposition ensembles for sales series (Gahirwal, 2013), and frameworks that forecast through the three latent GEV parameters or through joint wind speed, wind direction, and air pressure dynamics (Shang et al., 2020, Ambach et al., 2017). Among the provided sources, the title that explicitly uses the name is "TriForecaster: A Mixture of Experts Framework for Multi-Region Electric Load Forecasting with Tri-dimensional Specialization" (Zhu et al., 13 Aug 2025).

1. Scope and nomenclature

Within the provided sources, the term does not identify a single universally fixed architecture. Instead, it denotes forecasting designs whose structure is explicitly triadic: three branches, three stages, three latent parameters, or three specialization dimensions. This suggests that "TriForecaster" functions less as a canonical model family in the narrow sense than as a recurring design pattern for decomposing forecasting difficulty into three interacting subproblems.

Formulation	Three-way organization	Domain
Triformer (Cirstea et al., 2022)	triangular patch attention; variable-specific parameters	long sequence multivariate time series forecasting
T3Time (Chowdhury et al., 6 Aug 2025)	time, spectral, and prompt branches	multivariate time series forecasting
TriTS (Ao, 17 Apr 2026)	time, frequency, and 2D-vision spaces	long-term time series forecasting
Trio (Chen et al., 5 Jun 2026)	Temporal, Spatial, and Sample attention	multivariate time-series forecasting
TriForecaster (Zhu et al., 13 Aug 2025)	region, context, and time specialization	multi-region electric load forecasting

The statistical literature in the same corpus uses analogous three-part decompositions. Monthly sales forecasting is cast as trend, seasonality, and irregular decomposition (Gahirwal, 2013). Probabilistic weather post-processing is organized into model-specific Quantile Regression Forests, quantile averaging, and interpolation to a full predictive distribution (Kirkwood et al., 2020). Functional extremes are forecast by modeling the GEV location, scale, and shape parameters (Shang et al., 2020). Ternary probabilistic forecasting treats below-normal, near-normal, and above-normal outcomes as points inside a triangle, and couples visualization, verification, and recalibration in a single geometry (Jupp et al., 2011).

2. Long-sequence and multimodal neural formulations

A foundational long-sequence formulation appears in "Triformer: Triangular, Variable-Specific Attentions for Long Sequence Multivariate Time Series Forecasting--Full Version" (Cirstea et al., 2022). The model addresses two stated limitations of attention-based forecasters: canonical self attention has a quadratic complexity with respect to input length, and existing studies use the same model parameter space for all variables' time series. Triformer therefore proposes a triangular, variable-specific attention consisting of a novel patch attention with linear complexity, a triangular structure in which layer sizes shrink exponentially, and a light-weight method that enables distinct sets of model parameters for different variables' time series without compromising efficiency and memory usage (Cirstea et al., 2022). A representative formalization uses a multivariate time series $X \in \mathbb{R}^{T \times M}$ and predicts $\hat{Y} \in \mathbb{R}^{H \times M}$, with total complexity summarized as $O\!\left(\frac{w d}{p}\cdot T\right)$ under patching and triangular shrinking.

Later tri-modal architectures extend this logic from efficiency to representation disentanglement. T3Time defines a framework with time, spectral, and prompt branches, where the dedicated frequency encoding branch captures periodic structures and a gating mechanism learns prioritization between temporal and spectral features based on the prediction horizon (Chowdhury et al., 6 Aug 2025). The fusion is explicitly horizon-aware:

$$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$

The prompt branch uses a frozen GPT-2 and cross-modal alignment, and the paper reports an average reduction of $3.28\%$ in MSE and $2.29\%$ in MAE, with few-shot improvements of $4.13\%$ and $1.91\%$ in MSE and MAE with $5\%$ training data, and $3.62\%$ and $\hat{Y} \in \mathbb{R}^{H \times M}$0 with $\hat{Y} \in \mathbb{R}^{H \times M}$1 data (Chowdhury et al., 6 Aug 2025).

TriTS pushes tri-modality further by projecting $\hat{Y} \in \mathbb{R}^{H \times M}$2D time series into orthogonal time, frequency, and $\hat{Y} \in \mathbb{R}^{H \times M}$3D-vision spaces (Ao, 17 Apr 2026). Its Period-Aware Reshaping estimates a dominant period by

$\hat{Y} \in \mathbb{R}^{H \times M}$4

reshapes the normalized sequence into a temporal image, and processes it with Visual Mamba to model cross-period dependencies as global visual textures while maintaining linear computational complexity (Ao, 17 Apr 2026). The frequency branch uses Multi-Resolution Wavelet Mixing, the time branch retains a streaming linear anchor, and fusion is scale-aware:

$\hat{Y} \in \mathbb{R}^{H \times M}$5

$\hat{Y} \in \mathbb{R}^{H \times M}$6

On benchmark averages with lookback window $\hat{Y} \in \mathbb{R}^{H \times M}$7, TriTS records $\hat{Y} \in \mathbb{R}^{H \times M}$8 on ETTh1, $\hat{Y} \in \mathbb{R}^{H \times M}$9 on ETTh2, $O\!\left(\frac{w d}{p}\cdot T\right)$0 on ETTm1, $O\!\left(\frac{w d}{p}\cdot T\right)$1 on ETTm2, $O\!\left(\frac{w d}{p}\cdot T\right)$2 on Weather, $O\!\left(\frac{w d}{p}\cdot T\right)$3 on Electricity, and $O\!\left(\frac{w d}{p}\cdot T\right)$4 on Traffic, all as best averages in the reported table (Ao, 17 Apr 2026).

3. Retrieval, example organization, and structural priors

Trio redefines the triadic principle around attentional factorization rather than modality fusion. It is a sample-aware multivariate time-series forecasting architecture with Temporal attention for within-window dynamics, Spatial attention for inter-variable dependencies, and Sample attention for retrieval over historical lookback-future pairs (Chen et al., 5 Jun 2026). Context is reorganized into historical pairs and a current window with strict temporal ordering, with

$O\!\left(\frac{w d}{p}\cdot T\right)$5

so that leakage is prevented because the maximum index of each historical future segment is strictly smaller than the minimum index of the current window (Chen et al., 5 Jun 2026).

The key mechanism is Sample attention. Historical futures are embedded as prompts, current windows receive learnable future-query placeholders, and the encoder alternates Spatial, Temporal, and Sample attentions. The retrieval update is

$O\!\left(\frac{w d}{p}\cdot T\right)$6

followed by writing the updated future-query tokens back into the current window (Chen et al., 5 Jun 2026). This explicitly organizes long histories into examples rather than treating them as an undifferentiated monolithic sequence.

Trio is paired with a Time-Series Structural Causal Model generator that creates synthetic forecasting tasks with dynamic lags, delayed propagation, feedback loops, noise processes, and drift (Chen et al., 5 Jun 2026). The node update is

$O\!\left(\frac{w d}{p}\cdot T\right)$7

with dynamic lag scheduling

$O\!\left(\frac{w d}{p}\cdot T\right)$8

On synthetic delayed-dependency tasks, adding Sample Attention reduces MSE from $O\!\left(\frac{w d}{p}\cdot T\right)$9 to $$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$0 and MAE from $$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$1 to $$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$2 in Synthetic #1, and from $$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$3 to $$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$4 and from $$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$5 to $$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$6 in Synthetic #2 (Chen et al., 5 Jun 2026). On public benchmarks, Trio achieves the best average MSE on ETTm1 ($$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$7), ETTm2 ($$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$8), and Electricity ($$g = \sigma\!\left( W_4 \phi(W_3 g_{\mathrm{in}}) \right), \qquad Z_g = g \odot \tilde{F} + (1-g)\odot \tilde{Z}_t.$$9), while remaining competitive on Weather (Chen et al., 5 Jun 2026).

4. Statistical, probabilistic, and decomposition-based triads

In classical decomposition-based forecasting, the triadic structure appears as a decomposition of the signal itself. "Inter Time Series Sales Forecasting" models monthly sales by multiplicative decomposition,

$3.28\%$0

where the series is split into trend, seasonality, and irregular components (Gahirwal, 2013). Forecasting is performed componentwise with ARIMA/SARIMA experts, and the overall expert pool is the Cartesian product of $3.28\%$1 trend experts, $3.28\%$2 seasonal experts, and $3.28\%$3 irregular experts, yielding $3.28\%$4 triplets (Gahirwal, 2013). Consistent Model Mining with CHARM identifies experts that frequently co-occur in top APE ranks, and the method yields an average MAPE improvement of approximately $3.28\%$5 over Holt–Winters across $3.28\%$6 evaluated series (Gahirwal, 2013).

A tri-stage probabilistic formulation appears in weather post-processing. The framework in (Kirkwood et al., 2020) first learns model-specific forecast errors with Quantile Regression Forests, then combines per-model predictive quantiles by Vincentization,

$3.28\%$7

and finally interpolates the aggregated quantiles to a full predictive CDF (Kirkwood et al., 2020). In the road surface temperature application, $3.28\%$8 of observations fall within the $3.28\%$9 prediction interval overall, and the QRF post-processed median outperforms the median of available NWP forecasts across all lead times (Kirkwood et al., 2020).

For extreme values, the three components are latent distributional parameters. The functional extreme-value forecaster models the GEV location $2.29\%$0, scale $2.29\%$1, and shape $2.29\%$2 either as scalars or as smooth functions over an index (Shang et al., 2020). The functional parameterization is

$2.29\%$3

with coefficient vectors forecast by VAR models and uncertainty quantified by sieve bootstrap (Shang et al., 2020). In Monte Carlo experiments, the scalar formulation is best when parameters are truly scalar, while the functional formulation is best when $2.29\%$4 and $2.29\%$5, or all three parameters, are functional (Shang et al., 2020).

A different tri-variate statistical forecaster jointly predicts wind speed, wind direction, and air pressure through a threshold seasonal TVARX mean model and a threshold seasonal TARCHX variance model (Ambach et al., 2017). The six-dimensional response vector augments speed and pressure with sine/cosine Cartesian components of wind direction, and estimation is performed by iteratively re-weighted LASSO rather than ordinary likelihood estimation (Ambach et al., 2017). The paper reports accurate short-term point forecasts and prediction intervals up to twenty-four hours, with wind-direction yaw error remaining below $2.29\%$6 for up to about five hours ahead (Ambach et al., 2017).

Ternary probabilistic forecasting provides yet another triadic formalization. Forecasts $2.29\%$7 and observations are mapped into a triangle, with the Brier Score and Ranked Probability Score inducing associated geometries in which score, reliability, uncertainty, and resolution are interpreted as root-mean-square distances (Jupp et al., 2011). The decomposition is

$2.29\%$8

and the paper introduces a Ternary Reliability Diagram together with a nonlinear recalibration map on the simplex (Jupp et al., 2011).

5. TriForecaster for multi-region electric load forecasting

The explicit TriForecaster model of (Zhu et al., 13 Aug 2025) is a Mixture of Experts framework for Multi-Region Electric Load Forecasting under a Multi-Task Learning paradigm. Its central claim is that MRELF is governed by three challenges: regional variation, contextual variation, and temporal variation (Zhu et al., 13 Aug 2025). Regional variation arises from heterogeneous industrial structure and load composition; contextual variation from dependence on broader regimes such as weekday versus weekend or holiday versus regular day; temporal variation from changes in covariate impact across the day, week, and season (Zhu et al., 13 Aug 2025).

The architecture is correspondingly tri-dimensional. RegionMixer contains a shared expert and region-specific experts per layer, plus a stochastic fusion mechanism based on inter-expert distances rather than a task-specific gating network (Zhu et al., 13 Aug 2025). For region $2.29\%$9, the categorical distribution over shared and other region experts is

$4.13\%$0

and Stoch_Pool samples the per-position expert contribution (Zhu et al., 13 Aug 2025). CTSpecializer then applies ContextMoE along latent dimensions per timestamp and TimeMoE along the horizon dimension, again using stochastic fusion. A contrastive loss encourages context-coherent specialization through an InfoNCE-style objective based on positive and negative samples constructed from salient covariates (Zhu et al., 13 Aug 2025).

Experts in RegionMixer are implemented with TSMixer blocks, ContextMoE and TimeMoE use MLPs, and prediction is produced by a region-specific head $4.13\%$1 (Zhu et al., 13 Aug 2025). Training uses Adam with learning rate $4.13\%$2, batch size $4.13\%$3, latent dimension $4.13\%$4, and MSE plus the contrastive regularizer (Zhu et al., 13 Aug 2025). The datasets are EPC with three zone stations and $4.13\%$5, CEESC with two areas and $4.13\%$6, City-load with $4.13\%$7 cities and $4.13\%$8, and Bus-load with six buses and $4.13\%$9 (Zhu et al., 13 Aug 2025).

Empirically, TriForecaster yields an average MSE reduction of $1.91\%$0 and MAE reduction of $1.91\%$1 versus the second-best method across datasets (Zhu et al., 13 Aug 2025). Reported performance is $1.91\%$2 on EPC, $1.91\%$3 on CEESC, $1.91\%$4 on City-load, and $1.91\%$5 on Bus-load, all in MSE and MAE order (Zhu et al., 13 Aug 2025). Ablation on EPC shows degradation for every removed module: without RegionMixer, $1.91\%$6; without ContextMoE, $1.91\%$7; without TimeMoE, $1.91\%$8; full model, $1.91\%$9 (Zhu et al., 13 Aug 2025). The deployment on the eForecaster platform in eastern China covers $5\%$0 cities, supports a population exceeding $5\%$1 million and daily electricity usage over $5\%$2 gigawatt-hours, and is associated with $5\%$3 average accuracy improvement, about $5\%$4 million kWh saved per day, and about $5\%$5 USD daily cost savings (Zhu et al., 13 Aug 2025).

6. Comparative interpretation, misconceptions, and research directions

A common misconception is that TriForecaster necessarily denotes a transformer with three modalities. The provided literature does not support that restriction. Some variants are efficient attention models or multimodal neural networks (Cirstea et al., 2022, Chowdhury et al., 6 Aug 2025, Ao, 17 Apr 2026, Chen et al., 5 Jun 2026), but others are decomposition-based expert systems, probabilistic post-processing pipelines, VAR-on-parameters models, shrinkage-based multivariate statistical systems, or simplex-valued recalibration frameworks (Gahirwal, 2013, Kirkwood et al., 2020, Shang et al., 2020, Ambach et al., 2017, Jupp et al., 2011). A second misconception is that the triadic structure always refers to three input modalities; in fact, it may refer to three specialization axes, three forecast stages, three latent parameters, or three outcome categories.

Across these works, the triadic bias serves related technical purposes. It factorizes heterogeneity, separates long-range structure from local variation, or isolates calibration and uncertainty from deterministic prediction. This suggests that the shared principle is not the use of the number three as such, but the allocation of distinct forecast-relevant phenomena to distinct computational or statistical subspaces.

Open issues are stated differently across the corpus. TriTS identifies sensitivity to dominant-period estimation and proposes adaptive period detection and extension to multivariate irregular time series forecasting (Ao, 17 Apr 2026). Trio states that fully general PFN-style time-series forecasting remains an open problem, and that the realism of TS-SCM needs deeper validation (Chen et al., 5 Jun 2026). The electric-load TriForecaster identifies extensions to probabilistic forecasting, hierarchical reconciliation, and broader validation across provinces and countries (Zhu et al., 13 Aug 2025). The weather post-processing framework notes possible improvements from adaptive weighting and alternative learners for extreme lead times (Kirkwood et al., 2020). The extreme-value framework highlights the difficulty of functionally modeling $5\%$6 and points toward spatial functional extremes and regime-change handling (Shang et al., 2020).

Taken together, the literature presents TriForecaster as a recurrent research strategy for forecasting under entangled structure: separate the problem into three interacting components, assign each a dedicated inductive bias, and recombine them in a way that preserves efficiency, specialization, or calibration. In electric load forecasting this strategy appears as region, context, and time specialization (Zhu et al., 13 Aug 2025); in long-horizon multivariate forecasting as triangular attention and variable-specificity or tri-modal fusion (Cirstea et al., 2022, Chowdhury et al., 6 Aug 2025, Ao, 17 Apr 2026); and in probabilistic and statistical forecasting as three-stage post-processing, three-component decomposition, or three-parameter distributional modeling (Kirkwood et al., 2020, Gahirwal, 2013, Shang et al., 2020).