An Adaptive Multi-Scale Transformer for Green-Energy Stock Return Forecasting: A Controlled Evaluation of Static Scale Aggregation and Sample-Dependent Routing

Predicting green-energy stock returns requires distinguishing the benefits of model complexity from those of simpler forecasting rules. This study examines whether sample-dependent routing improves return forecasts over fixed or learned static aggregation of multiple temporal scales. A multi-scale Transformer framework combines daily and weekly price–volume inputs to forecast 5-, 10-, and 20-trading-day returns for a retrospectively assembled panel of 17 green-energy stocks. The evaluation includes financial baselines, repeated training runs, matched historical coverage, frozen-representation aggregation controls, and interventions on routing weights. Statistical comparisons account for both training randomness and temporal dependence, while constrained portfolio tests assess performance under transaction costs. Under the baseline training configuration in the primary test period, all evaluated neural configurations have higher mean squared errors, averaged across training runs, than the global training-mean predictor. Window-relative inputs improve forecasting performance, but adaptive routing does not show a consistent advantage over static aggregation. Apparent advantages supported by inference conditional on fitted models weaken when training variability is included. A revised training configuration improves pooled forecast accuracy relative to the historical mean in retrospective evaluations of earlier windows, but does not outperform the zero-return baseline. Portfolio results vary with input representation, forecast horizon, and allocation rule, without establishing a consistent economic advantage for adaptive routing. The study provides a controlled empirical assessment of scale aggregation and highlights the importance of input representation, simple financial benchmarks, and uncertainty across training runs. Its conclusions remain specific to the retrospective stock panel and evaluated periods.

Authors

Institutions

Publication Details

Journal
Computers
Published
2026-11-16
DOI
https://doi.org/10.3390/computers15110845
Primary Topic
Stock Market Forecasting Methods
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An Adaptive Multi-Scale Transformer for Green-Energy Stock Return Forecasting: A Controlled Evaluation of Static Scale Aggregation and Sample-Dependent Routing

Chenxue Yang, Xiyao Ma
Computers
Stock Market Forecasting Methods
article

An Adaptive Multi-Scale Transformer for Green-Energy Stock Return Forecasting: A Controlled Evaluation of Static Scale Aggregation and Sample-Dependent Routing

Chenxue Yang, Xiyao Ma
article en

Abstract

Predicting green-energy stock returns requires distinguishing the benefits of model complexity from those of simpler forecasting rules. This study examines whether sample-dependent routing improves return forecasts over fixed or learned static aggregation of multiple temporal scales. A multi-scale Transformer framework combines daily and weekly price–volume inputs to forecast 5-, 10-, and 20-trading-day returns for a retrospectively assembled panel of 17 green-energy stocks. The evaluation includes financial baselines, repeated training runs, matched historical coverage, frozen-representation aggregation controls, and interventions on routing weights. Statistical comparisons account for both training randomness and temporal dependence, while constrained portfolio tests assess performance under transaction costs. Under the baseline training configuration in the primary test period, all evaluated neural configurations have higher mean squared errors, averaged across training runs, than the global training-mean predictor. Window-relative inputs improve forecasting performance, but adaptive routing does not show a consistent advantage over static aggregation. Apparent advantages supported by inference conditional on fitted models weaken when training variability is included. A revised training configuration improves pooled forecast accuracy relative to the historical mean in retrospective evaluations of earlier windows, but does not outperform the zero-return baseline. Portfolio results vary with input representation, forecast horizon, and allocation rule, without establishing a consistent economic advantage for adaptive routing. The study provides a controlled empirical assessment of scale aggregation and highlights the importance of input representation, simple financial benchmarks, and uncertainty across training runs. Its conclusions remain specific to the retrospective stock panel and evaluated periods.

ComputersVol. 15(11)
National University of Singapore (SG), Agricultural Information Institute (CN), Chinese Academy of Agricultural Sciences (CN)
Stock Market Forecasting Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon