Raw Data vs Technical Indicators

A key design question in financial ML is:
Should we use raw market data, technical indicators like RSI, or a combination of both?

There is no universal answer — it depends on experimentation. While deep neural networks are known for their ability to automatically extract features from raw data, using hand-crafted indicators can sometimes still boost performance.

From a purist's perspective, relying on human-designed indicators might seem "unnatural," especially given the idea that neural networks should discover useful representations on their own. But in practice, what matters is performance, not philosophical elegance.

This dilemma resembles what happened in the development of AlphaGo:

• In early versions, human-designed features (e.g., liberties, influence maps) were used as input.
• In later versions like AlphaGo Zero, the model learned directly from raw board positions, without any human-engineered features.

Similarly, in trading, we may start with technical indicators — but our long-term goal is to train models that extract alpha directly from raw market signals, as data and compute scale.

Practical Tip

We recommend experimenting with all three:

1. Raw time-series data (prices, volume)
2. Common technical indicators (RSI, MACD, Bollinger Bands)
3. Hybrid input (raw + indicators)

... and evaluate their performance under your specific trading environment and reward function.