Decoding TINs: Reconstructing Classic Technical Analysis with Neural Networks

Background Recent advances combine deep learning with technical indicators for trading, yet most models lack domain‑specific architecture, dynamic adaptability, and interpretability. The authors propose Technical Indicator Networks (TINs) to systematically translate rule‑based indicators such as MACD, MA, and RSI into explainable, modular neural‑network structures.

Problem Definition

Algorithmic trading faces three core challenges: (1) absence of neural architectures designed for technical analysis, (2) fixed parameters of classic indicators limiting generalization across markets, and (3) reliance on single data sources (e.g., price) that miss cross‑market signals.

Method

3.1 Topological Representation of Indicators Each mathematical formula of a classic indicator is mapped to a network layer. For example, a moving‑average (MA) becomes a fully‑connected layer with uniform or exponentially decaying weights, forming a “fast‑MA → slow‑MA → subtraction → smoothing” hierarchy for MACD.

Similarly, the MACD calculation and its signal generation are expressed as successive layers, preserving the original mathematical relationships.

3.2 Network Construction and Training Strategy

Weight Initialization : Weights are set according to the exact formulas of indicators (e.g., SMA and EMA coefficients) so that the network reproduces classic indicator outputs before learning.

Training : Supervised learning uses historical trading signals as labels, while reinforcement learning (DDQN, PPO, Actor‑Critic) optimizes parameters in a simulated trading environment with a reward function reflecting profit and risk.

3.3 Key Operations and Modular Design

Subtraction Layer : Implements the fast‑MA minus slow‑MA operation of MACD.

Bias‑Regularized Division : Handles division in RSI, ROC, avoiding gradient instability.

Adaptive Pooling : 1‑D max/min pooling extracts high/low prices for stochastic oscillators.

Mean Absolute Deviation (MAD) : Computes price deviation for CCI.

3.4 Multi‑Dimensional Extension and Cross‑Market Analysis TINs accept heterogeneous inputs—price, volume, NLP‑derived news sentiment, macro indicators—and share the same topology across assets, enabling joint modeling of multiple stocks.

Experiments

4.1 Dataset and Setup Daily closing prices of the 30 US30 index constituents were used. Baselines: traditional MACD, TIN‑MACD with price only, TIN‑MACD with price + OBV, and a buy‑and‑hold US30 strategy. Evaluation metrics: Sharpe ratio, Sortino ratio, cumulative return. Training employed Deep Q‑Learning with a 52‑day price window and 26 hidden nodes mimicking MACD parameters.

4.2 Results

Risk‑adjusted returns: TIN‑MACD + OBV achieved Sharpe 2.7357 and Sortino 3.9886, far surpassing traditional MACD (1.6474 / 1.8450) and the buy‑and‑hold benchmark (1.4991 / 2.3174).

Cumulative return: 19.93 % vs. 14.16 % for traditional MACD.

Cross‑asset performance: In stocks such as NFLX, WMT, AAPL, TIN‑MACD consistently outperformed MACD in both Sharpe and Sortino.

Robustness: TIN‑MACD retains full interpretability—each node and connection corresponds to a financial formula—while reinforcement learning dynamically adapts parameters to varying market regimes.

Conclusion

TINs successfully bridge classic technical analysis and modern deep‑learning frameworks, delivering interpretable, adaptable, and higher‑performing trading strategies. By preserving indicator mathematics in network weights and leveraging reinforcement learning for dynamic optimization, TINs offer a promising path for AI‑driven quantitative finance.