Paper: arXiv 2507.20202

Authors: Longfei Lu

Abstract

Deep neural networks (DNNs) have transformed fields such as computer vision and natural language processing by employing architectures aligned with domain-specific structural patterns. In algorithmic trading, however, there remains a lack of architectures that directly incorporate the logic of traditional technical indicators. This study introduces Technical Indicator Networks (TINs), a structured neural design that reformulates rule-based financial heuristics into trainable and interpretable modules. The architecture preserves the core mathematical definitions of conventional indicators while extending them to multidimensional data and supporting optimization through diverse learning paradigms, including reinforcement learning. Analytical transformations such as averaging, clipping, and ratio computation are expressed as vectorized layer operators, enabling transparent network construction and principled initialization. This formulation retains the clarity and interpretability of classical strategies while allowing adaptive adjustment and data-driven refinement. As a proof of concept, the framework is validated on the Dow Jones Industrial Average constituents using a Moving Average Convergence Divergence (MACD) TIN. Empirical results demonstrate improved risk-adjusted performance relative to traditional indicator-based strategies. Overall, the findings suggest that TINs provide a generalizable foundation for interpretable, adaptive, and extensible learning architectures in structured decision-making domains and indicate substantial commercial potential for upgrading trading platforms with cross-market visibility and enhanced decision-support capabilities.

Complexity vs Empirical Score

  • Math Complexity: 6.0/10
  • Empirical Rigor: 4.5/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper introduces a novel neural architecture with moderate mathematical complexity involving vectorized operators and topological mapping, but the empirical validation is limited to a single-market backtest on DJIA constituents without detailed performance metrics or implementation details, leaning towards theoretical development over rigorous backtesting.

Research Flowchart

  flowchart TD
  A["Research Goal: Create an interpretable, adaptive neural architecture for algorithmic trading that modernizes classic technical indicators"] --> B["Methodology: Technical Indicator Networks (TINs)"]
  
  B --> C["Core Design: Reformulate rule-based indicator logic (e.g., MACD) into trainable neural modules with vectorized operators"]
  
  C --> D["Data/Inputs: Multi-dimensional market data from Dow Jones Industrial Average constituents"]
  
  D --> E["Computational Process: Apply analytical transformations (averaging, clipping, ratio computation) via modular layers, enable optimization through reinforcement learning"]
  
  E --> F["Key Findings: Improved risk-adjusted performance vs. traditional strategies, retains interpretability, demonstrates extensibility for adaptive trading systems"]
  
  F --> G["Outcomes: TINs provide a generalizable foundation for interpretable learning in structured decision-making domains with commercial potential for modernizing trading platforms"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes Technical Indicator Networks (TINs), a neural architecture that embeds the computational logic of traditional technical indicators and allows their parameters to be learned.

Source sentence
This paper introduces the Technical Indicator Networks (TINs) framework, which addresses this limitation by embedding the computational logic of technical indicators directly into neural network architectures.

Assumes 6

  • Transaction costs are ignored in the reported results. execution
    Source sentence
    Transaction costs were not considered, since this work is intended as a proof-of-concept demonstration.
  • The risk‑free rate is set to zero for Sharpe/Sortino calculations. model
    Source sentence
    Performance metrics are computed from daily portfolio returns, with the risk-free rate fixed at Rf = 0 for comparability.
  • All strategies are trained and tested on identical 52‑day rolling windows of historical adjusted closing prices. data
    Source sentence
    All strategies are trained and tested on identical 52-day rolling windows of historical adjusted closing prices.
  • Model outputs are discretized into binary buy or sell signals for position‑level simulation. execution
    Source sentence
    Model outputs are discretized into buy or sell signals for position-level simulation.
Show 2 more
  • A single hidden layer of 26 units initialized from canonical MACD parameters is sufficient for the proof‑of‑concept. model
    Source sentence
    The TIN architectures employ a single hidden layer of 26 units, initialized directly from the canonical MACD parameters (12, 26, 9) via mathematically grounded layer operators.
  • Trading frequency is low, with holding periods of several days to weeks. execution
    Source sentence
    Both strategies operate at relatively low trading frequencies—typically with holding periods of several days to weeks—yet their return distributions differ substantially.

Data

Backtest sample: 2023-08 – 2025-08 · US30 constituents · equities · daily

  • US30 daily adjusted closing prices · Yahoo Finance via yfinance library market-prices public
    2015-08-25 to 2025-08-22 · 30 constituents of the Dow Jones Industrial Average (US30)
    Source sentence
    The dataset comprises daily adjusted closing prices for the 30 constituents of the US30 index, retrieved via the publicly available yfinance library. The sample is divided into an in-sample training period from 2015-08-25 to 2023-08-21 and an out-of-sample evaluation period from 2023-08-22 to 2025-08-22.

Recognised sources: Yahoo Finance

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
Transaction costs were not considered, since this work is intended as a proof-of-concept demonstration.

Finds 3

Reported performance (TIN-MACD (Price+OBV)): Sharpe 3.53 (benchmark 1.12) · total return 23.1%

  • Both TIN‑MACD variants outperform the buy‑and‑hold US30 index on risk‑adjusted metrics and cumulative returns. performance
    Source sentence
    both TIN‑MACD configurations (price-only and OBV-augmented) achieve higher cumulative returns than the canonical MACD, while also outperforming the buy-and-hold baseline of the US30 index in terms of risk-adjusted measures.
  • Statistical tests show the price+OBV TIN‑MACD Sharpe improvement (+0.3575) is significant (p = 0.0202), while the price‑only variant is not (p = 0.2113). statistical
    Source sentence
    TIN‑Price+OBV variant achieves a substantially larger mean improvement of +0.3575, which is statistically significant at the 5% level (p = 0.0202).
  • Sortino improvements are also significant for the OBV‑augmented TIN‑MACD (+1.0910, p = 0.0485) and borderline for the price‑only variant (+1.0371, p = 0.0573). statistical
    Source sentence
    the TIN‑Price+OBV variant achieves a mean increase of +1.0910, which is statistically significant (p = 0.0485).

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (27)

  1. Improving MACD Technical Analysis by Optimizing Parameters and Modifying Trading Rules : Evidence from the Japanese Nikkei 225 Futures Market (2021) — link
  2. Deep learning with long short-term memory networks for financial market predictions (2018) — link
  3. A deep learning framework for financial time series using stacked autoencoders and long short-term memory (2017) — link
  4. Predicting Stock Market Trends Using Machine Learning and Deep Learning: The Case of Tehran Stock Exchange (2020) — link
  5. An ensemble deep learning model for stock price prediction (2021) — link
  6. Feature selection and deep neural networks for stock price prediction (2021) — link
  7. A Deep Learning Approach for Forecasting Cryptocurrencies Using ARIMA-LSTM (2021) — link
  8. DRL-based portfolio management with transaction cost (2022) — link
and 19 more
  1. A robust time-series validation approach for financial forecasting with deep neural networks (2022) — link
  2. Stock price prediction with optimized deep LSTM network (2023) — link
  3. Predictability of stock returns using neural networks (2023) — link
  4. Real-time forecasting of time series in financial markets (2023) — link
  5. Stock price prediction for new energy vehicle companies using deep learning approaches (2024) — link
  6. Interpretable stock price forecasting model using genetic algorithms and machine learning (2023) — link
  7. Scientometric review and analysis of recent approaches to financial forecasting (2023) — link
  8. An application of deep reinforcement learning to algorithmic trading (2021) — link
  9. Outperforming algorithmic trading reinforcement learning baselines (2022) — link
  10. Algorithmic trading using continuous action space deep reinforcement learning (2024) — link
  11. An Advisor Neural Network framework using LSTM-based informative stock analysis (2025) — link
  12. Financial applications of machine learning: A literature review (2023) — link
  13. Multi-scale contrast approach for stock index prediction (2025) — link
  14. Key technical indicators for stock market prediction (2025) — link
  15. Deep Direct Reinforcement Learning for Financial Signal Representation and Trading (2017) — link
  16. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning (1992) — link
  17. High-Dimensional Continuous Control Using Generalized Advantage Estimation (2016) — link
  18. Asynchronous Methods for Deep Reinforcement Learning (2016) — link
  19. Human-level control through deep reinforcement learning (2015) — link

0 of 27 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.