TRACE-CASH: Trial-History-Conditioned Reinforcement Learning for Adaptive Configuration Exploration in Time-Series CASH

Yu-Han Huang; Yujia Wu; Vincent S. Tseng

Source-linked AI summary

TRACE-CASH: Trial-History-Conditioned Reinforcement Learning for Adaptive Configuration Exploration in Time-Series CASH

Yu-Han Huang, Yujia Wu, Vincent S. Tseng

arXiv:2608.16410v1Aug 17, 2026cs.LG

TL;DR

Controlled comparisons of heterogeneous search methods in time-series CASH remain limited because existing studies use differing spaces and evaluation protocols. This paper introduces TRACE-CASH, a trial-history-conditioned hybrid optimizer, and finds the complete procedure has the lowest mean rank on MASE and WQL across 41 task variants.

Problem
Controlled comparisons of heterogeneous forecasting search methods under aligned time-series CASH spaces and evaluation protocols remain limited.
Method
TRACE-CASH combines trial-history-conditioned grouped candidate generation with fixed rules for model coverage, validation-guided exploitation, and recovery.
Results
Across 41 task variants, TRACE-CASH has the lowest mean rank on MASE and WQL and the lowest descriptive window-averaged test-MASE rank in full and late windows.
Takeaways & Limitations
TRACE-CASH is competitive among the evaluated TS-CASH methods under aligned evaluation conditions.
Takeaways & Limitations
SLLMBO* is a source-based adaptation rather than an official reproduction.

Abstract

from arXiv · show

Combined algorithm selection and hyperparameter optimization (CASH) searches a conditional space in which the selected model determines which hyperparameters are active. In time-series forecasting, temporal choices, chronological validation, and costly evaluations further complicate this search. Controlled comparisons of heterogeneous search methods under a shared time-series CASH (TS-CASH) evaluation protocol remain limited. Within this setting, we study TRACECASH, a task-local hybrid sequential optimizer combining grouped actor-critic candidate generation with fixed rules for model coverage, validation-guided exploitation, and exploration after stalled progress. A model actor proposes an initial forecasting model; three model-conditioned actors generate temporal, architectural, and training actions; and a modelspecific decoder constructs the configuration ultimately evaluated. We compare TRACE-CASH with six alternatives spanning random, Bayesian, evolutionary, multi-objective, and language-model-assisted search across 41 dataset-frequency task variants. TRACE-CASH has the lowest mean rank on both MASE and WQL. Descriptively, it also has the lowest window-averaged test-MASE rank in the predefined full and late windows. These results support the complete TRACECASH procedure as competitive among the evaluated methods.

Introduction

Forecasting heterogeneity and conditional model-dependent choices motivate AutoML for time-series forecasting, while chronological validation and costly evaluations make TS-CASH comparisons difficult. TRACE-CASH addresses this through a controlled comparison of seven heterogeneous methods under aligned evaluation conditions and a task-local hybrid search procedure.

Motivation: Forecasting tasks vary in domain, sampling frequency, series length, and missing-data patterns, so no single model is uniformly best.This heterogeneity motivates AutoML systems that automate temporal feature construction, neural architecture choices, and pipeline hyperparameters.
Evaluation challenge: Chronological validation and computationally expensive deep-model evaluations require aligned temporal protocols and finite trial budgets.
Comparison problem: Existing forecasting AutoML studies differ in candidate spaces, optimized components, objectives, and evaluation protocols, confounding comparisons of search strategies with study-specific design.TRACE-CASH uses a common nine-family conditional search-space design for controlled comparison.
TRACE-CASH: TRACE-CASH is a task-local sequential optimizer reinitialized for each task–seed run and updated from accumulating within-task trial history.It combines grouped actor–critic candidate generation with fixed rules for model coverage, validation-guided exploitation, and recovery.
Study design: The study compares seven heterogeneous methods using the same nine-family conditional search space, chronological evaluator, attempted-trial budget, validity checks, and incumbent rule.The complete TRACE-CASH procedure is assessed across 41 task variants.

Related Work

Related work spans forecasting benchmarks, time-series AutoML, conditional CASH, decision decomposition, and language-model-assisted search. TRACE-CASH is positioned within these lines by combining sequential conditional configuration search with grouped decision-making for forecasting CASH.

Forecasting and HPO benchmarks: Forecasting benchmarks include the Monash Archive, TFB, YAHPO Gym, and TSBench, which provide diverse datasets, unified evaluation, repeatable surrogate scenarios, or offline HPO comparisons.TSBench records 97,200 evaluations across 20 datasets and compares Random Search, HyperBand, SMAC, and BOHB offline.
Time-series AutoML: Forecasting AutoML automates data preparation, model and hyperparameter selection, pipeline construction, and ensembling through systems such as AutoAI-TS, AutoGluon–TimeSeries, and auto-sktime.The cited systems respectively rank pipelines from multiple model classes, ensemble probabilistic forecasters, or use Bayesian optimization, warm starts, and multi-fidelity evaluation.
Conditional CASH: CASH jointly selects a learning algorithm and its hyperparameters in a hierarchical conditional space where only hyperparameters associated with the selected algorithm are active.SMAC3 and TPE address mixed, conditional, or tree-structured configuration spaces; forecasting must accommodate model-specific temporal choices.
Decision decomposition: Decision-decomposition methods distribute optimization among specialized agents, as in MA2ML and MAT-HPO; TRACE-CASH applies related grouping to conditional forecasting CASH.In TRACE-CASH, the evaluated model determines which hyperparameters are active.
Language-model-assisted search: Language-model-assisted optimizers use language models for warm starts, sampling, surrogate modeling, proposer routing, or suggestions combined with TPE.The cited methods are LLAMBO, LB-MCTS, and SLLMBO.

Problem Formulation

The problem formulation defines a shared nine-family conditional forecasting search space and evaluates candidate configurations under a finite trial budget with chronological validation and testing. Configurations are ranked by validation MASE among valid trials, with earliest-trial tie breaking and no incumbent when none are valid.

Search Space: All methods search a common nine-family conditional space, where each model family m∈M activates its own hyperparameter domain Θ_m.The formulation represents the space using a finite set of forecasting-model families M and model-specific domains Θ_m.
Candidate Configurations: Each candidate is λ=(m, θ_m)∈Λ, but admissibility does not guarantee successful fitting or forecasting.A candidate combines one model family with hyperparameters from that family’s domain.
Evaluation Budget: Under budget B, methods attempt candidates λ_t∈Λ for t=1,…,B, and each attempt consumes budget regardless of outcome.Candidates may be generated conditionally on earlier trial outcomes.
Trial Eligibility: Validation MASE ranks only valid trials that complete successfully, produce finite scores, and pass shared forecast-output validity checks.Test-stage validity also remains part of shared eligibility checks.
Selection Rule: The incumbent is the earliest attempt attaining minimum validation MASE; if no valid trial exists, no incumbent is defined.Numerical test scores neither rank configurations nor break validation-MASE ties.

Proposed Method: TRACE-CASH

TRACE-CASH is a hybrid sequential optimizer that combines history-conditioned learned candidate generation with fixed rules for model coverage, validation-guided exploitation, and recovery. Its grouped actors and model-specific decoder respect the conditional TS-CASH structure while adapting decisions from within-task trial history.

Core procedure: TRACE-CASH combines learned, history-conditioned candidate generation with fixed rules for model coverage, validation-guided exploitation, and recovery.The learned policies operate within the same search procedure as fixed coverage and recovery rules.
Grouped conditional generation: Four actors share each trial-history state: one proposes the model, while three model-conditioned actors generate temporal, architectural, and training actions.The decoder applies only hyperparameters relevant to the model ultimately selected for evaluation.
Task-local adaptation: All learned components are reinitialized for every task–seed run, while fixed rules provide initial coverage, guide later allocation using outcomes, and reopen exploration after stagnation.These rules remain fixed across task–seed runs and complement learned candidate generation.
Trial-history state: The state summarizes recent outcomes, search progress, best validation MASE, model coverage, and model-use history, excluding numerical test-score magnitudes.Each trial updates the replay buffer and next state, so later decisions depend on accumulated within-task history.
Learning and evaluation: Valid trials receive clipped negative validation MASE rewards, whereas failed or invalid trials receive fixed penalties; rewards train the actor–critic components without defining incumbent eligibility or ranking.Critics learn from evaluated configurations, while actor updates sample new actions without treating them as evaluated trials.

Experimental Setup

The study compares seven search methods under aligned TS-CASH evaluation conditions across 41 dataset–frequency task variants. It uses fixed chronological evaluation and validation-based incumbent selection while preserving method-specific candidate-generation mechanisms and objectives.

Common evaluation protocol: Seven search methods share a nine-family conditional search space, chronological evaluation, attempted-trial budgets, validity checks, and a validation-MASE incumbent rule.Candidate generation, priors, initialization, and internal objectives remain method-specific.
Tasks and models: 41 dataset–frequency task variants span 13 application domains, 10 normalized sampling frequencies, varied horizons, series counts, and lengths.The model pool comprises DLinear, DeepAR, DeepNPTS, WaveNet, TiDE, SimpleFeedForward, TFT, PatchTST, and LagTST.
Compared methods: The six comparators are Random Search, SMAC3, TPE, NSGA-III, MOTPE, and SLLMBO*, covering random, Bayesian, evolutionary, multi-objective, and language-model-assisted strategies.SLLMBO* is a source-based adaptation combining language-model and TPE candidate generation with trial-history context.
Optimization objectives: NSGA-III and MOTPE optimize validation MASE, training time, and parameter count, but every reported incumbent is selected solely by validation MASE.SLLMBO* adapts prompts and active hyperparameter spaces to TS-CASH rather than reproducing the official method.
Evaluation and reporting: Test scores neither rank eligible configurations nor break ties, WQL evaluates the validation-selected incumbent, and test-MASE trajectories are summarized by mean rank over early, late, and full windows.Runs use fixed chronological validation and testing routes, with undefined budget entries before an incumbent exists.

Results

TRACE-CASH achieves the lowest mean rank on both MASE and WQL, including a post-hoc sensitivity analysis. Its descriptive budget profiles also show the lowest mean rank in the predefined late and full windows, evaluating the complete integrated procedure.

Full-budget results: TRACE-CASH has the lowest mean rank on both reported metrics, MASE and WQL.WQL evaluates probabilistic quality for the validation-MASE-selected incumbent rather than a WQL-selected optimum.
Sensitivity analysis: TRACE-CASH retains the lowest mean rank on both MASE and WQL across 29 source-group blocks in a post-hoc sensitivity analysis.
Budget profiles: TRACE-CASH has the lowest mean rank in the predefined late and full windows of the descriptive budget profiles.Table 3 summarizes trajectories using window-averaged rank.
Procedure-level interpretation: The seven-method evaluation characterizes TRACE-CASH as an integrated search procedure combining learned candidate generation, exploration, and fixed search rules.

Conclusion

TRACE-CASH combines trial-history-conditioned grouped candidate generation with fixed rules for model coverage, validation-guided exploitation, and recovery. Across 41 task variants, it achieves the lowest mean rank on MASE and WQL and the lowest descriptive window-averaged test-MASE rank in the predefined full and late windows.

Conclusion: TRACE-CASH combines trial-history-conditioned grouped candidate generation with fixed model-coverage, validation-guided exploitation, and recovery rules.The complete procedure integrates learned candidate generation with fixed search-control rules.
Conclusion: Across 41 task variants, TRACE-CASH has the lowest mean rank on MASE and WQL.This ordering is retained in a separate post-hoc sensitivity analysis using 29 source-group blocks.
Conclusion: TRACE-CASH has the lowest descriptive window-averaged test-MASE rank in the predefined full and late windows.The result is descriptive rather than presented as an additional inferential comparison.