Multi-Horizon Prediction of Financial Technical Indicators via Ticker-Aware Deep Learning and Ensemble Optimization

Ahmed Ayoubi () and Arnaud Vaissaire
Additional contact information
Ahmed Ayoubi: UFRM - UCA - UFR de Mathématiques - Clermont-Auvergne - UCA - Université Clermont Auvergne
Arnaud Vaissaire: Meiren

Working Papers from HAL

Abstract: We study the joint prediction of the continuous future values of two widely used technical indicators-the Moving Average Convergence Divergence (MACD) and the Stochastic Oscillator-across 15 simultaneous horizons (h = 1 to 15, i.e., 5 to 75 minutes ahead) on 15 heterogeneous assets (forex, cryptocurrencies, equities) using 5-minute intraday data from 2010-2025 (53.7M raw rows; 4.0M after quality filtering). Predicting indicator values directly, rather than binary signals or raw prices, is operationally motivated: systematic trading systems consume indicator values as inputs to signal generators, EMA-based indicators give smoother regression targets than raw returns, and bounded oscillators such as the Stochastic avoid the non-stationarity of raw prices. To our knowledge, this continuous multi-output multi-horizon formulation across heterogeneous assets has not been explicitly addressed in prior work, which mostly targets binary signals or single-asset, single-horizon settings. Our contributions are: (i) a comparison of 12 architectures, from ARIMA and Ridge to XGBoost, LightGBM, CatBoost, Bi-LSTM, GRU, TCN, and TFT; (ii) per-asset target normalization and a learned ticker embedding, associated with a 2.8× increase in MACD Pearson correlation over a global LSTM baseline (0.265 to 0.753); (iii) an ablation study across three axes; and (iv) walk-forward validation over three non-overlapping windows suggesting stable behavior of GRU v3 (std ≈ 0.004). The GRU+TFT ensemble reaches 64.3%/63.5% Directional Accuracy (val/test) on MACD, versus 0% for the persistence baseline, which has higher Pearson correlation (0.773 vs. 0.718) but never predicts a directional change. For the Stochastic Oscillator, DA reaches 69.9%/69.8% (val/test), consistent across all evaluated learned models.

Keywords: Computing methodologies → Neural networks; CCS Concepts; Ablation study; Ensemble optimization; Intraday forex; Ticker embedding; Temporal Fusion Transformer; GRU; Stochastic oscillator; MACD; Technical analysis; Computer systems organization → Recurrent neural networks Multi-horizon forecasting; CCS Concepts Computing methodologies → Neural networks (search for similar items in EconPapers)
Date: 2026-10
Note: View the original document on HAL open archive server: https://hal.science/hal-05800405v1
References: Add references at CitEc
Citations:

Downloads: (external link)
https://hal.science/hal-05800405v1/document (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:hal:wpaper:hal-05800405

DOI: 10.5281/zenodo.23378336

Access Statistics for this paper

More papers in Working Papers from HAL
Bibliographic data for series maintained by CCSD ().