TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting on GIFT-Eval, Selected Entirely on the Training Split

Nathan Thierry; Andre-Louis Rochet

cs.AISep 16, 2026

TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting on GIFT-Eval, Selected Entirely on the Training Split

Authors: Nathan Thierry, Andre-Louis Rochet

Organizations: TW3 Partners

Abstract

TW3Cast is a time-series forecasting system that reaches position 3 of 130 entries on the GIFT-Eval benchmark by mean MASE rank, as of 2026-09-14. The two entries above it belong to the leaderboard's agentic category, multi-step systems that use agents or language models to reason about, generate or select forecasts. TW3Cast runs no agent and no language model. Its selection is a table computed once on the training split and then frozen, and its experts are public foundation models lightly fine-tuned on those training splits. For each of the 97 dataset, frequency and horizon configurations, the table serves one of four modes: a specialist, which is a LoRA or full fine-tune of Chronos-2, TiRex or Toto whose training data was cleaned and enriched by explicit rules; a quantile blend that contains a specialist; a blend of base models; or a selection tournament played on a backtest carved from the training split. Every decision in the table was taken on that backtest. A specialist is admitted the moment it beats the tournament there, so a candidate costs a few megabytes and minutes of GPU time, and a failed candidate changes nothing. Three guarded mechanisms protect the selection from its own biases: a dual accuracy and calibration criterion, an asymmetric margin against candidates that saw the series during training, and conservative per-window gates. The selection rules themselves were chosen inside a temporal meta-backtest. The best base model served alone reaches a mean MASE rank of 33.8, the tournament served on every configuration reaches 38.0, and the full router reaches 19.4. The routing table, the expert index, the pinned base-model revisions, the submitted score file and the dated snapshot of the public scores are released, and every leaderboard number in this paper regenerates from them by one script.

Figures & tables

Panel of Figure 1 — Figure 1: The frozen router. A configuration name resolves by table lookup to a designated specialist, a quantile blend or the tournament mode. Per-window gates then bound the tail risk, and the system outputs the nine deciles for each evaluation window.

Figure 2: Composition of the released router. Left, the serving mode of the 97 configurations by horizon class. Right, the architecture family of the 43 designated specialists. Both panels are read from router.csv and experts.json by make_figures.py .

Panel of Figure 3 — Figure 3: Every decision is taken on a backtest extracted from the training split. The remaining training data trains the experts, and the resulting router is frozen before any test window is read.

Panel of Figure 4 — Figure 4: The expert pipeline. Cleaning and dosed fusion, controlled windows with robust normalization and pinball loss, three architectures with cross blends and conditional scaling, then the admission rule.

Position	Entry	Mean MASE rank	Declared type	Test leakage
1	STRIDE_w_Synapse	14.9	agentic	No
2	EXAONE-Forecast-Agent	19.2	agentic	No
3	TW3Cast (this work)	19.4	router	No
4	LS-MoE	19.7	agentic	No
5	LS-Agent	21.7	agentic	No
6	Falcon-Agent	22.5	agentic	No

Table 1: Top of the GIFT-Eval leaderboard by mean per-configuration MASE rank on 2026-09-14, computed by make_figures.py from the per-configuration result files of the 129 public entries with complete results, with the submitted TW3Cast result file inserted as one more entry and ranked by the rule of Section 8 . Declared type and leakage are copied from each entry’s own description file on the leaderboard. The snapshot of the public files, our result file and the script are in the release.

Figure 5: Mean per-configuration MASE rank of the 130 entries of Table 1 , sorted by position, with the six members of the tournament pool and TW3Cast marked; the number in parentheses is the position. Script make_figures.py .

Pool member	Position	MASE rank	WQL rank	Declared type
Chronos-2	44	47.2	47.6	pretrained
TiRex	60	58.4	51.1	zero-shot
TiRex-2	36	44.2	42.7	pretrained
Toto 2.0 2.5B FT	23	33.8	33.1	fine-tuned
TimesFM 2.5	57	53.5	53.1	zero-shot
TurkForecast LoRA	45	48.1	47.2	fine-tuned

Table 2: Standing of each tournament pool member as its own public leaderboard entry, in the same field and with the same rank computation as Table 1 . The WQL column is the same average on weighted quantile loss. TiRex-2 is listed under its pretrained entry, the one used in the pool.

Figure 6: Per-configuration MASE of TW3Cast relative to the best pool member on that configuration, in percent, over the 97 configurations sorted by the ratio. Negative bars are configurations where the system beats the per-configuration oracle over the pool. Script make_figures.py .

Variant	All 97	Tournament configurations (27)
Chronos-2	47.2	42.6
TiRex	58.4	63.2
TiRex-2	44.2	41.5
Toto 2.0 2.5B FT	33.8	28.1
TimesFM 2.5	53.5	58.8
TurkForecast LoRA	48.1	51.4

Table 3: Mean per-configuration MASE rank of each variant in the field of Table 1 , over all 97 configurations and over the 27 configurations that the released router serves by tournament. Pool members are their public entries and keep the ranks of Table 2 . The tournament-only variant is scored from its released per-configuration grid, inserted into the same field as one more entry; on the tournament configurations that grid matches the submission to within 0.2 percent. Script make_figures.py .

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Mechanism	Constant	Value	Set by
Anti-memorization margin	$\delta$	12%	meta-backtest
Near-tie CRPS tie-break	gap	$<3\%$	meta-backtest
Blend fallback	gap to winner	$<5\%$	meta-backtest
Per-window gate	$\tau$ (foundation / naive)	1.5 / 3.0	meta-backtest
Per-window gate	challenger advantage	$\geq 10\%$	meta-backtest
Anti-catastrophe gate	$\tau$	5 to 10	meta-backtest

Appendix

Table 4: All fixed constants of the released system. Every constant was chosen inside the train-side protocols of Sections 5 and 7 and then frozen.

Configuration	Served by
bitbrains_fast_storage/5T/long	tournament
bitbrains_fast_storage/5T/medium	E01 (Toto)
bitbrains_fast_storage/5T/short	tournament
bitbrains_fast_storage/H/short	E02 (Toto)
bitbrains_rnd/5T/long	Toto + C2 + Turk
bitbrains_rnd/5T/medium	Toto + C2 + Turk

Appendix

Table 5: The released routing table, first half of the 97 configurations of router.csv in alphabetical order, with the members of each blend. Members named E nn are designated specialists, with their architecture family in parentheses; other names are base models of the tournament pool at their pinned revisions; blend members carry equal weights. Generated by make_figures.py .

Configuration	Served by
jena_weather/10T/long	tournament
jena_weather/10T/medium	E44 (C2)
jena_weather/10T/short	tournament
jena_weather/D/short	E20 (TiRex) + TFM
jena_weather/H/long	tournament
jena_weather/H/medium	E43 (C2) + TFM + C2

Appendix

Table 6: The released routing table, second half, same conventions as Table 5 .

Explore similar work

Jun 10, 2026cs.LG

TimeRouter: Efficient and Adaptive Routing of Time-Series Foundation Models

Time-series foundation models (TSFMs) are increasingly explored as predictive experts within emerging agentic time-series systems. However, TSFMs exhibit heterogeneous inductive biases, and no single model consistently dominates across forecasting regimes, making expert selection a critical challenge. Existing systems often delegate this decision to LLM-based controllers, incurring substantial inference overhead. We present TimeRouter, an efficient routing framework that leverages empirical complementarity across a pool of pretrained TSFMs through lightweight discriminative routing, selective gating, and ensemble fallback. Concretely, TimeRouter combines a learned routing head, a selective gate, and an ensemble fallback, enabling adaptive expert selection without invoking an LLM at inference time. TimeRouter achieves state-of-the-art performance on the GIFT-EVAL leaderboard, with an LB MASE of 0.6765. Beyond benchmark performance, our ablation studies provide empirical insights into TSFM routing design, highlighting the importance of pool composition and selective gating. Taken together, these results position TimeRouter as a modular and lightweight routing layer for future agentic time-series systems built upon foundation-model pools. Our code is available at https://github.com/UConn-DSIS/TimeRouter.

Kanghui Ning, Yushan Jiang, Kashif Rasul +3

School of Computing, University of Connecticut, Storrs, USA · Department of Machine Learning Research, Morgan Stanley, New York, USA.

Adaptive Model Routing Time Series Forecasting Time-Series Foundation Models

Ask PDF

Sep 23, 2026cs.LG

Forecast Workflow Bench: Evaluating Language-Model Decisions with Budgeted Forecast Tools

Time-series foundation models (TSFMs) provide forecasts for operational decisions, but accuracy alone does not determine their value. Evaluating agents that use these models requires measuring decision quality and forecast cost. FWBench evaluates this capability on 1,251 electricity and cycle-hire cases using fixed forecast tools and simulated capacity contracts. Agents select models, histories and horizons, then submit capacities to minimize a stated loss-cost objective. We evaluated two hosted and eight local configurations, including small language models, and tested local models with and without TSFMs. GPT-6 Astra bought inexpensive short-horizon forecasts selectively, using 2.5% of the budget, and outperformed fixed policies when the saved decisions were scored with three loss-cost weightings. FWBench enables reproducible evaluation of how language models select and use time-series forecasts to make decisions under cost constraints.

Shunya Nagashima

Neurogica Inc.

Forecasting Benchmarks LLM Tool Use Time-Series Foundation Models

Ask PDF

Apr 13, 2026cs.LG

TempusBench: An Evaluation Framework for Time-Series Forecasting

Foundation models have transformed natural language processing and computer vision, and a rapidly growing literature on time-series foundation models (TSFMs) seeks to replicate this success in forecasting. While recent open-source models demonstrate the promise of TSFMs, the field lacks a comprehensive and community-accepted model evaluation framework. We see at least four major issues impeding progress on the development of such a framework. First, existing evaluation frameworks comprise benchmark forecasting tasks derived from often outdated datasets (e.g., M3), many of which lack clear metadata and overlap with the corpora used to pre-train TSFMs. Second, these frameworks evaluate models along a narrowly defined set of benchmark forecasting tasks, such as forecast horizon length or domain, but overlook core statistical properties such as non-stationarity and seasonality. Third, domain-specific models (e.g., XGBoost) are often compared unfairly, as existing frameworks do not enforce a systematic and consistent hyperparameter tuning convention for all models. Fourth, visualization tools for interpreting comparative performance are lacking. To address these issues, we introduce TempusBench, an open-source evaluation framework for TSFMs. TempusBench consists of 1) new datasets which are not included in existing TSFM pretraining corpora, 2) a set of novel benchmark tasks that go beyond existing ones, 3) a model evaluation pipeline with a standardized hyperparameter tuning protocol, and 4) a tensorboard-based visualization interface. We provide access to our code on GitHub: https://github.com/Smlcrm/TempusBench and maintain a live leaderboard at https://smlcrm.com/tempusbench.

Denizalp Goktas, Gerardo Riaño-Briceño, Omkar Tekawade +14

Simulacrum New York City, NY, USA

Time Series Forecasting Forecasting Benchmarks Time-Series Foundation Models

Ask PDF