Systematic trading borrows most of its methods from statistics and machine learning. What makes financial data unusual is that it is noisy, non-stationary, and adversarial: any pattern you find is being competed away as you find it.
That context matters more than any individual technique, and it is why methods that dominate other fields often disappoint here.
The statistics of data ordered in time, covering autocorrelation, stationarity and volatility clustering. Price series are non-stationary, which is why models are usually built on returns rather than prices, and why volatility models such as GARCH exist to capture the tendency of turbulent periods to cluster together.
Turning raw market data into inputs a model can use: returns over various horizons, volatility measures, volume ratios, order book imbalance. In practice this determines model quality far more than the choice of algorithm, and it is also where look-ahead bias most often creeps in.
Classifying the market into states such as trending, ranging or high-volatility, often using clustering or hidden Markov models. The value is in switching strategies rather than predicting price: knowing which environment you are in is more reliable than knowing where price goes next.
Training an agent to learn a trading policy by rewarding outcomes. It is appealing because it optimises the actual objective rather than a proxy, and it is difficult in practice because financial data is limited, non-stationary and noisy, so agents frequently learn the history rather than a general policy.
Inputs beyond price and volume: on-chain activity, social data, funding, positioning. The attraction is that fewer people use it. The reality is that it is often shorter in history, less clean, and harder to validate, so its edge has to be weighed against the difficulty of testing it honestly.
The platform is not a research environment, and pretending otherwise would waste your time. It is the execution and validation layer around whatever research you do.
Start in paper mode, size up when it earns it. Nothing to install.
