[Springer 2021] Deep Learning for Investment: Clustering Indicators and Residual Blocks
Deep learning-based investment strategy: technical indicator clustering and residual blocks
This paper introduces three deep learning frameworks (CNN-TIC, EO-CNN-TIC, and ResNet-TA) for financial trend forecasting. By converting 1-D time-series technical indicators into 2-D multi-channel "images" and employing residual learning and evolutionary optimization, the proposed ResNet-TA model achieves a record-breaking 36.70% annual return on DOW-30 stocks.
TL;DR
Predicting the stock market is often dismissed as a "random walk," but this study proves that sophisticated data representation—turning indicators into multi-channel images—combined with ResNet architectures and Cuckoo Search optimization can yield annual returns of over 36%, far exceeding the benchmark SOTA and "Buy and Hold" strategies.
Background & Motivation: Moving Beyond 1D Signals
While LSTMs and RNNs are the traditional choice for time-series, they often struggle with the "spatial" relationships between different technical indicators (e.g., how the RSI interacts with Moving Averages). The authors observe that:
- Data Imbalance: Financial markets are mostly in a "Hold" state. Traditional models often ignore rare but profitable "Buy/Sell" triggers.
- Exit Strategy: Deep learning models usually predict a direction but ignore the execution (when to take profit).
The core insight here is to treat technical indicators as a cohesive "eye" on the market, grouping them by behavior to let a CNN "see" the market's health.
Methodology: The Three Pillars of Performance
1. Technical Indicator Clustering (CNN-TIC)
Instead of a flat 1-D vector, the authors use 15 indicators across 15 time intervals (6 to 20 days), creating a 15x15 image. They categorize these into 5 clusters:
- Cluster 1: Oscillators (RSI, Williams %R)
- Cluster 2: Trend/Averages (SMA, EMA, WMA, etc.)
- Cluster 3/4: Momentum and Volume indicators.
- Cluster 5: Directional/Money flow indicators.
This creates a 5-channel image, allowing the CNN to learn features within each indicator type before merging them in higher layers.
2. Cuckoo Search Optimization (EO-CNN-TIC)
Trading isn't just about the signal; it's about the exit. The authors used Cuckoo Search, a nature-inspired metaheuristic, to find the "sweet spot" for risk management. The optimal values found were a 20% Take Profit and a 1% Stop Loss, which maximized the fitness function (Cumulative Profit).
3. Going Deeper with ResNet-TA
To solve the accuracy degradation found in deep LeNet-style models, the authors implemented Residual Blocks. By adding shortcut connections, they allowed the model to be deeper (extracting more abstract features) without losing the identity of the original input.
Results: Crushing the Market Benchmarks
The results across 10 years (2007–2016) of Dow Jones 30 and ETF data were remarkable:
- Annual Returns: ResNet-TA reached 36.70%, compared to just 10.47% for a Buy-and-Hold strategy.
- Recall: Use of class-weighted loss (giving 8x more importance to "Sell" signals) pushed the recall to over 0.90, meaning the model almost never misses a significant market move.
Critical Insight & Conclusion
The success of ResNet-TA highlights two critical lessons for AI in finance:
- Spatial Representation Matters: Transforming 1D signals into a 2D clustered image allows CNNs to utilize their inductive bias for pattern recognition effectively.
- Execution is Part of the Model: Integrating evolutionary optimization for exit triggers (Take Profit/Stop Loss) bridges the gap between a "classifier" and a "trading system."
Limitations: While the returns are high, the volatility (Standard Deviation) of ResNet-TA on individual stocks is also high (13.16%), suggesting that while highly profitable, the model can have aggressive drawdowns on specific tickers. Future research incorporating LSTMs to handle the temporal flow of these "images" could be the next frontier.