Foundation Forecasting Models (FFMs)

• Foundation Forecasting Models (FFMs) are large, pretrained time-series models that learn reusable temporal patterns from vast, diverse data.
• They leverage transformer architectures, spatiotemporal inductive biases, and multi-modal fusion to produce accurate and probabilistic forecasts.
• FFMs enable zero-shot, few-shot, and fine-tuned adaptation across domains like clean energy, finance, and Earth-system prediction while addressing operational constraints.

Foundation Forecasting Models (FFMs) are large, pretrained time-series models that learn general-purpose temporal representations from massive, diverse data and are then adapted, in zero-shot, few-shot, or fine-tuned form, to downstream forecasting tasks. In current usage, the term covers both generalist time-series foundation models and domain-specialized forecasters for clean energy, retail demand, finance, building systems, epidemic surveillance, power systems, weather, and Earth-system prediction. Across these settings, FFMs are distinguished by broad pretraining, transferability across domains and horizons, long-context sequence modeling, and an increasing emphasis on probabilistic outputs, spatiotemporal inductive biases, and operational constraints such as calibration, latency, and privacy (Ferdaus et al., 30 Jul 2025).

1. Definition, scope, and taxonomy

The most explicit general definition in the recent literature describes FFMs as large, pretrained models for time series that learn reusable temporal structure from broad corpora and are subsequently adapted to downstream forecasting. In clean energy forecasting, this definition is sharpened to models explicitly designed or adapted for renewable generation and demand forecasting, with domain-specific spatiotemporal inductive biases, multi-modal fusion of weather and sensor data, and risk-aware probabilistic outputs. In demand forecasting, the same term is instantiated as pretrained transformer backbones adapted to hierarchical retail prediction; in building energy systems it denotes pretrained, probabilistic transformer forecasters; in finance it is often used for large pretrained time-series models such as TimesFM, Chronos, Moirai, and Time-MoE; and in Earth-system science it denotes pretrained models that learn joint spatiotemporal statistics across heterogeneous physical variables and can be specialized through finetuning (Ferdaus et al., 30 Jul 2025).

A common taxonomy organizes FFMs along four axes.

Axis	Main categories	Illustrative instances
Architecture	transformer-only, spatial/graph-aware, hybrid vision-sequence, diffusion/generative, mixture-of-experts	TimesFM, GridFM, Aurora, Time-MoE
Pretraining paradigm	generative next-step/sequence, masked reconstruction, contrastive/self-supervised, multi-task	autoregressive, MAE-style, InfoNCE
Adaptation	zero-shot, few-shot, full fine-tuning, PEFT, domain adaptation, continual learning	LoRA, adapters, BitFit, prompt tuning
Application domain	wind, solar PV, electricity load, retail demand, finance, buildings, epidemics, Earth system	shared backbone with domain specialization

This taxonomy also reveals an important ambiguity in the term. Some papers use FFM to mean a broad, cross-domain generalist model; others use it for a domain-specialized system built on foundation-model principles. This suggests that “foundation” in forecasting refers less to a single architecture than to a training-and-transfer regime: large-scale pretraining, reuse across tasks, and adaptation under data or deployment constraints.

2. Architectural patterns and inductive biases

Transformer-only sequence models remain the dominant architectural family. The clean-energy review highlights decoder-only transformers such as Lag-Llama, TimesFM, and the decoder extension of TimeGPT-1, alongside encoder–decoder models. Patching reduces sequence length by grouping consecutive time steps into tokens, improving efficiency and accuracy on long horizons, while frequency-aware branches add spectral components to capture diurnal and weekly cycles. The standard attention core remains

$$A = \operatorname{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$

For wind and solar forecasting, spatiotemporal embeddings incorporate time and geolocation, with additional learned context such as terrain, plant layout, and turbine height; graph-aware variants such as GridFM represent the grid as nodes and edges and combine graph operations with temporal transformers to capture flow and congestion dependencies (Ferdaus et al., 30 Jul 2025).

Multi-modal fusion is a second recurrent pattern. Solar forecasting systems combine satellite or sky-camera imagery with irradiance or PV-output time series; weather-oriented models such as Aurora use a 3D Swin Transformer over atmospheric cubes together with Perceiver IO for multi-resolution, multi-variable ingestion. In Earth-system forecasting, ESFM generalizes this idea through variable-wise tokenization, axial attention over the variable axis, NaN tokens for missing data, and multi-resolution tokenizers that let a single backbone ingest dense reanalysis grids, climate-model outputs, sparse satellite swaths, and station observations. ESFM also introduces adaptive layer-norm-based ensembles to convert a deterministic backbone into a probabilistic forecaster without separate backbone passes (Ozdemir et al., 20 Apr 2026).

Mixture and routing mechanisms appear in several forms. Billion-scale sparse-expert designs such as Time-MoE and Moirai-MoE specialize experts to regimes such as coastal wind or continental solar. SEMPO replaces heavy sparse MoE with a lightweight Mixture-of-PrOmpts enabled Transformer, in which token-dependent routing selects prompt experts, while an energy-aware spectral decomposition module separates high-energy and low-energy frequency components before patchification. With 6.5M parameters and pretraining on only approximately 83M time points, SEMPO is presented as a lightweight foundation model rather than a large-scale one, yet it retains zero-shot and few-shot behavior across heterogeneous datasets (He et al., 22 Oct 2025).

A further architectural tendency is explicit adaptation or refinement around a frozen base model. LoRA-style low-rank updates insert

$$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$

into attention or MLP blocks to adapt a general model to local plants, climates, or buildings while limiting parameter updates (Ferdaus et al., 30 Jul 2025). In finance, RefineBridge instead leaves the base TSFM frozen and learns a complementary generative refinement module based on a Schrödinger Bridge objective, treating TSFM forecasts as priors and learning stochastic transport toward observed targets (Bolton et al., 25 Dec 2025).

3. Pretraining, transfer, and personalization

The pretraining objectives used in FFMs follow a relatively stable repertoire. Generative forecasting optimizes next-step or next-segment prediction; masked time-series modeling reconstructs masked segments in MAE style; contrastive learning aligns augmentations across scales or modalities through InfoNCE; and multi-objective training shares supervision across variables and horizons. In clean-energy forecasting, these objectives are often combined with synthetic rare-event augmentation and physics-guided features; in SEMPO, pretraining is an explicit two-stage procedure combining energy-aware self-supervised reconstruction and prompt tuning with forecasting-plus-reconstruction losses (Ferdaus et al., 30 Jul 2025).

Adaptation spans a wide range of regimes. Zero-shot use is prominent in generalist models such as TimesFM, Chronos, and Chronos-Bolt, especially when rapid deployment or data scarcity makes local retraining costly. Few-shot and full fine-tuning remain central in domains where local dynamics are idiosyncratic or physically constrained. The literature repeatedly reports that PEFT methods—LoRA, adapters, BitFit, and prompt tuning—can preserve pretrained priors while substantially reducing adaptation cost. In building energy systems, LoRA applied to Chronos attention projections reduced training FLOPS by 33% and improved per-iteration speed by 2.3× relative to full fine-tuning, while maintaining or improving forecast quality across occupancy, CO2, plug loads, and HVAC signals (Park et al., 31 May 2025).

Continual and online learning are increasingly treated as part of the FFM paradigm rather than as post hoc engineering. The clean-energy review notes light periodic updates to track seasonal changes, new assets, EV adoption, and weather shifts. Financial realized-volatility forecasting makes this explicit: TimesFM zero-shot is described as a reasonable baseline, but incremental fine-tuning is reported as essential for learning volatility patterns effectively. The best average MSE in that study was achieved by $\mathrm{TFM64}^{\log}_{IL} = 0.00115$, while $\mathrm{TFM512}_{PT}$ achieved the best QLIKE of $0.19253$, outperforming HAR at approximately $0.20083$ (Goel et al., 16 May 2025).

Personalization under privacy constraints has also entered the design space. Time-FFM proposes a federated foundation model in which time series are transformed into LM-compatible token sequences, a prompt adaptation module dynamically selects domain-customized prompts, a GPT-2 backbone is frozen, and a shared global encoder is trained by federated aggregation while local prediction heads remain client-specific. Its federated objective couples global encoder updates with local heads, and in the reported experiments it achieves an overall average MSE/MAE of $0.555/0.434$, outperforming several federated baselines and remaining competitive with centralized cross-domain models (2405.14252).

4. Forecast forms, uncertainty, and coherence

A central conceptual development is the distinction between forecast accuracy and forecast form. Recent work classifies foundation-model forecasts into point, quantile, parametric, and trajectory-ensemble outputs, and argues that the form constrains the downstream tasks the model can support. Trajectory ensembles are strictly more expressive because they can be marginalized into points, quantiles, or parametric summaries without extra assumptions, whereas converting marginal forecasts into coherent trajectories requires explicit dependence modeling through copulas or conformal methods. The same work proves that marginals alone cannot determine path-dependent event probabilities, since distinct joint distributions can share identical marginals while assigning different probabilities to threshold crossings, sums, or maxima (Perez-Diaz et al., 22 Oct 2025).

This distinction matters directly for operational forecasting. Point forecasts target central tendency; quantile and parametric forecasts support marginal uncertainty; trajectory ensembles are required for pathwise bands, first-passage times, window aggregates, and scenario generation. The paper maps six canonical tasks to minimal sufficient forecast types and recommends evaluation aligned with the target task: MAE or MSE for deterministic accuracy, pinball loss or CRPS for marginal probabilistic forecasts, Energy Score and Variogram Score for joint trajectories, and Brier-type metrics for event probabilities (Perez-Diaz et al., 22 Oct 2025).

Energy and power-system applications add a second structural requirement: coherence across horizons and hierarchies. FFMs commonly produce intra-hour, day-ahead, and week-ahead outputs through multi-head or hierarchical decoders, and then reconcile them with penalties that enforce consistency between aggregated short-horizon and long-horizon forecasts. Spatial and hierarchical reconciliation, including MinT-style projection, is used to impose coherence from asset to region to system level. The clean-energy review also emphasizes calibrated probabilistic outputs through Gaussian NLL, quantile loss, CRPS, isotonic or Platt scaling, deep ensembles, Monte Carlo dropout, conformal prediction, and diffusion-based scenario generation (Ferdaus et al., 30 Jul 2025).

Conformal prediction has become an especially important post hoc layer for zero-shot FFMs. In a direct comparison between TSFMs and traditional models under split conformal prediction, the foundation-model setting is reported to have two advantages: better point forecasts when training data are scarce, and more stable calibration because little or no data are spent on training. On ERCOT with only 2,232 points available and a 24-hour horizon, Chronos-Bolt achieved MASE $0.362$, MSIW $0.312$, and MCR $91.3\%$, while comparable LightGBM splits produced larger interval widths and weaker point accuracy (Achour et al., 9 Jul 2025).

5. Domain-specific instantiations and empirical evidence

The empirical record for FFMs is highly domain-dependent. In clean energy, the review synthesizes a broad set of gains: Chronos is credited with 15–20% error reduction in solar forecasting; Lag-Llama with approximately 8% improvement in wind prediction and probabilistic outputs; TimesFM with near fully-trained accuracy in zero-shot across tasks; Aurora with comparable weather accuracy at approximately 30% faster than NWP; and contrastive curriculum with an improvement of about 14.6% versus naive fine-tuning for building energy few-shot performance (Ferdaus et al., 30 Jul 2025).

Retail demand forecasting provides a different pattern: rather than relying on a single backbone, performance improves through structured ensembling around foundation models. The dual-strategy framework of hierarchical ensemble (HE) and architectural ensemble (AE) partitions forecasting by semantic retail level and combines heterogeneous backbones including LightGBM, DNN, DeepAR, PatchTST, TEMPO, and Chronos. On M5, HE improved PatchTST from WRMSSE $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$0 to $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$1, TEMPO from $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$2 to $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$3, and Chronos from $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$4 to $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$5; AE further reduced average WRMSSE to $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$6 for LightGBM + PatchTST (Yang et al., 29 Jul 2025).

In commercial building load forecasting, the contrast between zero-shot and fine-tuned FFMs is especially sharp. Fine-tuned TimesFM on the ComStock-based IL-HET subset achieved $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$7 NMSE/NMAE at $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$8, compared with $$W' = W + AB,\quad A \in \mathbb{R}^{d\times r},\; B \in \mathbb{R}^{r\times d},\; r \ll d,$$9 for zero-shot TimesFM and $\mathrm{TFM64}^{\log}_{IL} = 0.00115$0 for the best base model, TimesNet. On IL-HOM, fine-tuned TimesFM achieved $\mathrm{TFM64}^{\log}_{IL} = 0.00115$1 at $\mathrm{TFM64}^{\log}_{IL} = 0.00115$2, again substantially outperforming zero-shot. The authors conclude that dataset heterogeneity and architecture matter more than parameter count, while also reporting that zero-shot TimesFM is not competitive in this setting (Bose et al., 2024). A probabilistic study on a real commercial net-zero building reached a similar conclusion: zero-shot TSFMs were generally suboptimal, whereas fine-tuned Chronos with PEFT improved Occupancy from MASE $\mathrm{TFM64}^{\log}_{IL} = 0.00115$3 to $\mathrm{TFM64}^{\log}_{IL} = 0.00115$4, wQL from $\mathrm{TFM64}^{\log}_{IL} = 0.00115$5 to $\mathrm{TFM64}^{\log}_{IL} = 0.00115$6, and MSIS from $\mathrm{TFM64}^{\log}_{IL} = 0.00115$7 to $\mathrm{TFM64}^{\log}_{IL} = 0.00115$8 (Park et al., 31 May 2025).

Financial forecasting likewise resists simple zero-shot narratives. In realized-volatility forecasting, TimesFM zero-shot is reasonable but insufficient; incremental fine-tuning is repeatedly reported as necessary for strong performance, with DM and GW tests favoring fine-tuned variants over classical econometric baselines (Goel et al., 16 May 2025). In broader financial forecasting, a later line of work argues that standard LoRA often degrades TSFMs on non-stationary, heavy-tailed series, whereas the RefineBridge post-processing module improves the base model in 81 of 90 configurations and reduces MSE by 11%–71% across assets and horizons. For example, on the S&P 500 at $\mathrm{TFM64}^{\log}_{IL} = 0.00115$9, Chronos moved from original MSE $\mathrm{TFM512}_{PT}$0 to RefineBridge $\mathrm{TFM512}_{PT}$1, while LoRA degraded to $\mathrm{TFM512}_{PT}$2 (Bolton et al., 25 Dec 2025).

Power systems sit between these extremes. A benchmark on ERCOT reports that pure zero-shot FFMs underperform operational thresholds, with TimesFM the best among FFMs but still at roughly 9.3% nMAE for solar, 8.7% for wind, and 7.8% for load in day-ahead forecasting. After fine-tuning, however, TimesFM and Chronos-Bolt reach day-ahead nMAE of approximately 3.7–3.9% for solar and wind, while TimesFM achieves approximately 2.82% for load. The same benchmark reports that Chronos-Bolt yields the lowest CRPS for load and wind, and that FFMs degrade less than LSTM or CNN on unseen sites (Za'ter et al., 23 Apr 2026).

Epidemic and Earth-system forecasting broaden the scope further. In regional influenza forecasting, pretrained numeric transformers and a mixture-of-experts fusion model outperform LLM-style time-series methods; the best overall model, MultiFoundationCore, reaches average MSE $\mathrm{TFM512}_{PT}$3 and NNSE $\mathrm{TFM512}_{PT}$4 for 1–4 week ILI forecasting, while hospitalization-pretrained PatchTST and iTransformer variants outperform their from-scratch counterparts (Jafari et al., 17 Jun 2026). In the Earth system, ESFM demonstrates that the foundation-model paradigm can be extended beyond scalar sequences to heterogeneous spatiotemporal fields, achieving 6-hour MAE values such as $\mathrm{TFM512}_{PT}$5 K for T850 and $\mathrm{TFM512}_{PT}$6 K for T2m, competitive CRPS in an 8-member ensemble, stable multidecadal rollouts, and strong performance on sparse MODIS PWV and station data (Ozdemir et al., 20 Apr 2026). UniExtreme narrows the domain further to extreme weather forecasting, reporting approximately 10.9% improvement in extreme-region MAE and 9.85% in RMSE over the strongest baseline while reducing the MAE Gap between extreme and general regimes (Ni et al., 2 Aug 2025).

6. Operational constraints, controversies, and research agenda

The literature is increasingly explicit that FFMs are not uniformly superior to specialist models. A realistic operational study across traffic, ETTh1, exchange rates, and M4 Daily shows that zero-shot FFMs perform best where periodic structure transfers readily and where cold-start or long-tail deployment makes local training impractical, but specialists retain an edge in physically constrained systems and in stringent latency or governance settings. On Traffic, TimesFM 2.0 achieves the best MASE at $\mathrm{TFM512}_{PT}$7, but XGBoost is best on sMAPE and RMSE; on ETTh1, XGBoost dominates with MASE $\mathrm{TFM512}_{PT}$8; on Exchange, TimesFM 2.5 reduces MASE to $\mathrm{TFM512}_{PT}$9, well below XGBoost’s $0.19253$0, yet XGBoost remains best on sMAPE and RMSE. The same study reports large latency differences, from $0.19253$1 ms P95 for XGBoost to $0.19253$2 ms for TimesFM 2.0 and $0.19253$3 ms for Chronos, and proposes a Complexity Router that routes only about 30% of series to FFMs, yielding a 70% cost reduction versus universal FM deployment while improving MASE versus universal specialists (Soni et al., 23 May 2026).

A second controversy concerns benchmarking itself. A realistic evaluation of cross-frequency transfer learning argues that many FFM gains are overstated by small evaluation suites, improper averaging across datasets of radically different sizes, weak statistical baselines, and leakage between pretraining and test corpora. In that study, carefully configured statistical models and their ensembles outperform existing FFMs by more than 8.2% in sCRPS and by more than 20% in MASE across 15 public competition datasets, even though synthetic pretraining still improves an FFM by about 7% (Olivares et al., 23 Sep 2025). This does not negate the FFM paradigm, but it shifts emphasis from scaling alone to data curation, leakage control, and evaluation protocol.

Resource constraints form a third fault line. The hospitality benchmark reports that Chronos-Bolt matched ML models in 3 out of 4 restaurants with zero-shot inference and minimal feature engineering, but also notes that CPU-only execution was suboptimal and that ML meta-models remained the top performers overall in that hourly 14-day scenario (Arab et al., 5 Feb 2025). Lightweight alternatives such as SEMPO, and privacy-preserving architectures such as Time-FFM, respond directly to this tension by reducing model size or avoiding raw-data centralization (He et al., 22 Oct 2025).

Several open directions recur across domains. One is more faithful domain structure: clean-energy work stresses physics-guided constraints, graph structure, and coherent uncertainty; epidemic forecasting highlights mechanistic alignment between pretraining and target dynamics; Earth-system models emphasize heterogeneous data ingestion and variable-wise coupling (Ferdaus et al., 30 Jul 2025). A second is better output design: the forecast-form literature suggests that many operational failures attributed to “accuracy” are actually failures to produce the right object—trajectory ensembles rather than isolated marginals (Perez-Diaz et al., 22 Oct 2025). A third is trustworthy deployment: interpretability, calibration, regulatory compliance, and formal testing remain central, especially in critical infrastructure and finance.

Taken together, the evidence suggests a qualified synthesis. FFMs are most compelling when they combine large-scale pretraining with domain-appropriate inductive biases, efficient adaptation, coherent uncertainty, and task-aligned evaluation. A plausible implication is that the mature form of the field will not be a single universal forecaster replacing all specialists, but an ecosystem of generalist backbones, lightweight or federated adapters, structured ensembles, and routing mechanisms that allocate forecasting problems to the model class best matched to their data regime and operational objective.