Deep Learning Applications in Hierarchical Time Series Forecasting of Market Value
Deep learning models RNN, LSTM and GRU applied to hierarchical DJIA forecasting boost accuracy when reconciled. MinT (WLS-struct) beats Bottom-Up and Top-Down, cutting errors up to 15% over 115- and 230-day horizons. Study integrates reconciliation techniques and compares AvgRMSE, AvgMAPE, AvgRelRMSE and AvgRelMAPE. Findings stress forecast coherence and weighting. Neural architectures gain unevenly from methods. Training spanned market, sector and stock levels. Limitations include lack of out-of-sample tests beyond DJIA and sensitivity to data frequency.
What it examines
This study applies hierarchical time series forecasting to the DJIA market value at three levels (total index, sectors, individual stocks). It uses deep learning models (RNN, LSTM, GRU) for base predictions and evaluates coherence through reconciliation methods (Bottom-Up, Top-Down, Middle-Out, MinT, Optimal Combination).
What it concludes
Reconciliation substantially improves forecast accuracy, with MinT (WLS-struct) performing best across short and long horizons for all models. This approach aids investors, analysts, and policymakers in coherent market value forecasts. Future work may explore additional deep architectures, alternative reconciliation techniques, and broader financial markets applications.
Evidence objects
MinT (WLS-struct) reconciliation with RNN, LSTM and GRU for hierarchical DJIA forecasting model comparisons yield cuts errors up to 15% outperforming Bottom-Up and Top-Down across both 115-day and 230-day horizons.
key_findings bullet 1 · key_findings · validation V0
Authors integrate advanced neural networks with multiple reconciliation techniques, comparing performance via four metrics (AvgRMSE, AvgMAPE, AvgRelRMSE, AvgRelMAPE) through rigorous evaluation to underscore forecast coherence and structural weighting in modeling.
key_findings bullet 2 · key_findings · validation V0
Models trained on DJIA data at market, sector and stock levels reconciled for coherence; results reveal varying benefits among architectures and reconciliation noting limitations in out-of-sample validation and frequency sensitivity.
key_findings bullet 3 · key_findings · validation V0
Applying RNN, LSTM and GRU alongside hierarchical reconciliation (MinT, bottom-up, top-down), this paper uniquely integrates deep learning and coherent multi-level forecasting for DJIA values. Although methods are established, its comparative analysis and synthesis deliver fresh, unique insights and practical applications for quantitative finance, informing robust, reconciled market predictions with impact.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … , financial analysts, and policymakers due to the inherent volatility and complexity of financial … To capture sequential dependencies in hierarchical financial data, deep …
Source row: 530 · abstract type: snippet