Document Type
Conference Material
Abstract/Description
Equal-weighted ensembling is often treated as a default upgrade over individual machine learning models. We test this assumption on one-day-ahead stock-price prediction across five sector diversified U.S. assets (2015β2024) by comparing Random Forest (RF), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and their equal-weighted average under identical features, preprocessing, and chronological splits. Pooled across assets, LSTM attains the lowest test-set RMSE (17.70), SVR is statistically tied with LSTM under the Diebold and Mariano test (π = 0.494), the equal-weighted ensemble is third (18.13), and RF is last (20.33); SVR beats the ensemble at π = 0.001 (Bonferroni-significant) while LSTMbeats it at π = 0.035 (marginal under family-wise correction). The ensemble inherits intermediate accuracy and a positive bias between those of its members. The bias and variance decomposition predicts this ranking when constituent models have unequal accuracy, correlated errors, and heterogeneous bias; from the reported error standard deviations the implied average pairwise error correlation is approximately 0.94, and the heterogeneous-bias and unequal-accuracy conditions are visible directly. Equal-weighted ensembling does not automatically improve on careful individual model selection in this regime; validating its underlying assumptions on validation data is the methodological lesson.
Department
Computer Science and Information Systems
First Page
519
Last Page
524
DOI
https://doi.org/10.1145/3838457.3838535
Date
9-2026
Conference Title
IoTML 2026: 2026 6th International Conference on Internet of Things and Machine Learning
Conference Location
Nanning China
Citation Information
Khadka, Shraddha; Thapa, Prashidda; Sharma, Pragyan; and Silwal, Shreya, "Do Ensemble Models Always Win? A Comparative Machine Learning Evaluation for Financial Time-Series Prediction" (2026). Student Publications. 4.
https://lair.etamu.edu/cose-student-publications/4
