How to build a systematic strategy

A systematic strategy is an investment process written as rules. It does not mean the strategy is simple, fully automated, or free from judgment. It means the important choices are made before trading starts: what data is used, how signals are built, how the model is tested, and how the portfolio is sized.

The best way to understand quant trading is not to start with a model. Start with the workflow. Every serious strategy, whether it is a simplemomentum rule or a large hedge fund platform, moves through the same chain: data, features, fit, portfolio, execution, review.

The workflow in one page

1. Data collection: define what the strategy can know

The first question is not "what model should we use?" It is "what does the strategy know at decision time?" That includes prices, volume, fundamentals, analyst estimates, news, sector membership, macro data, or alternative data. A strategy using yesterday's closing prices is very different from one using live order-book data.

Good data collection is boring in the best possible way. You want complete histories, consistent symbols, corporate actions handled correctly, and no future information leaking into the past. If the data is wrong, a sophisticated model only becomes a faster way to learn the wrong lesson.

2. Feature engineering: turn raw data into useful signals

Raw data is rarely the thing you trade. You usually transform it into features: a 12-month momentum score, a short-term reversal score, a volatility estimate, a valuation ratio, a liquidity measure, or a sector-relative earnings trend.

There are many valid ways to build features. You can use different horizons, normalize within sectors, combine slow and fast signals, cap outliers, or make features conditional on the market regime. This is where domain knowledge matters. A feature is not just a number; it is a market hypothesis translated into code.

3. Fit and validation: avoid fooling yourself

The fitting step asks whether the features actually predict future returns. Sometimes this means a simple ranking rule. Sometimes it means a regression, tree model, neural network, or ensemble. The method is less important than the discipline around testing.

The danger is overfitting: building a strategy that explains the past perfectly because it memorized random noise. A serious process uses a train/test split, walk-forward testing, simple benchmarks, and realistic backtests that include costs.

4. Portfolio optimization: make the signal tradable

A signal says what looks attractive. A portfolio says how much capital to put behind each idea. Portfolio optimization balances expected return, volatility, correlation, turnover, liquidity, sector exposure, and leverage.

This is where many beginner strategies break. A signal can be real but too expensive to trade. A model can rank stocks well but accidentally load the whole portfolio into one sector. An optimizer can reduce risk but also dilute the edge if its constraints are too tight.

The main choices to tune

A useful mental model

Think of a systematic strategy as a factory. Data is the raw material. Features are the processed parts. The model is the quality-control system. Portfolio optimization is the assembly line. Execution is delivery. If any stage is weak, the finished product is weak, even if the other stages look impressive.