An Ensemble Deep Learning Approach for Forecasting Top 20 TSX-Listed Equities

Abstract

Forecasting equity prices on the Toronto Stock Exchange (TSX) is a genuinely difficult problem, and not just for the usual reasons. The Canadian market sits at the crossroads of two forces that rarely move together: global commodity cycles, which drive the resource sector that makes up a large share of the index, and tightly regulated domestic banking, which provides the steady counterweight. When oil prices fall sharply, the banks often hold firm; when interest rates spike, the energy producers can still benefit from high commodity prices. This push-and-pull makes the TSX a more complex forecasting environment than markets that are dominated by a single sector or a handful of mega-cap technology companies.

This study takes that complexity head-on by building and rigorously evaluating an ensemble deep learning pipeline against ten years of daily price data — January 2015 through January 2025 — for twenty of the TSX's most prominent listed equities. The tickers were chosen to span four distinct sectors: banking and financial services, energy and infrastructure, technology and precious metals, and diversified industrials. This breadth is deliberate; a model that only works within one industry has limited practical value.

Adjusted closing prices were pulled from the Yahoo Finance API, cleaned for gaps and corporate actions, transformed into log-returns to achieve approximate stationarity, and enriched with rolling statistical features before any modelling began. The resulting architecture combines two LSTM sub-networks — each trained independently on a 60-day sliding window of return features — with weights initialised from a checkpoint pretrained on approximately 500 S&P 500 constituent series. Predictions from the two sub-networks are aggregated through a compact dense regression head to produce a single one-step-ahead price forecast.

The proposed model outperforms the Li-Shi Hybrid Preprocessing Neural Network model [3] on all three evaluation metrics: RMSE falls to 0.02195 (a 4.6% reduction), MAE to 0.01594 (11.4% reduction), and MAPE to 0.38% (15.6% reduction). More importantly, residual diagnostics confirm that forecast errors are unbiased and homoscedastic across the full range of predicted values — a property that matters enormously if these forecasts are to be used in a real risk management context, where reliable confidence intervals are not optional. The full pipeline is open-source and fully reproducible from a standard Python environment.