Open research

Creating a systematic strategy step by step, in the open

Open research · Published 2026-10-23

Most failed trading strategies fail for the same reasons: an idea without a mechanism, data that flatter the past, a backtest tuned until it looks good, and no plan for what happens when reality differs. A disciplined workflow cannot guarantee a profitable strategy, but it can stop you fooling yourself. This article lays out the stages of creating a systematic strategy, as practised in our open research series.

Stage 1: a hypothesis with a mechanism

Start with a reason why a pattern should exist, and why it should persist. Good hypotheses usually fit one of a few categories: a risk premium (investors are paid for bearing a risk others avoid), a behavioural bias (investors systematically misprice something), a structural constraint (some participants must trade regardless of price, such as index funds at rebalancing), or an informational advantage. Momentum, for instance, is supported by both behavioural and risk explanations, as our momentum overview describes. A pattern found by searching data without a mechanism is much more likely to be noise.

Stage 2: specify the rules in writing

Before touching data, write the rules down precisely: the universe of instruments, the signal and its look-back, the entry and exit conditions, the position sizing, the rebalancing schedule and the cost assumptions. Commit this to the project repository with a date. Every later change is then visible as a deliberate decision rather than a quiet adjustment.

Stage 3: source and clean the data

• Use prices adjusted for splits and dividends, and know which adjustment method your source applies.
• Include delisted securities if the universe is individual stocks, to avoid survivorship bias.
• Use point-in-time data for fundamentals: a company's earnings are known only after they are reported, and later restatements must not leak backwards.
• Check for gaps, duplicate dates, stale prices and extreme outliers, and document how each is handled.

Stage 4: an honest first backtest

Run the rules as written on a defined in-sample period. Make sure signals are computed using only information available before the trade; a common safeguard is to compute signals at one day's close and trade at the next day's open or close. Include commissions, bid-ask spreads and an estimate of market impact. Record the result whatever it is.

Report more than the return: volatility, Sharpe ratio, maximum drawdown and its duration, turnover, the number of trades, the worst months and the skewness of returns. A strategy with an attractive average and a hidden crash risk, like momentum in rebounds, will only show up in these details.

Stage 5: decompose the returns

Regress the strategy's excess returns on known factors, such as the market, size, value, profitability, investment and momentum, as described in our article on the Fama-French five-factor model. If the factors explain the returns, the strategy may still be useful, but it is a way of buying known premiums rather than a new source of return. Knowing that affects how it should be combined with other strategies.

Stage 6: test robustness

• Parameter sensitivity: vary each parameter around its chosen value. A robust strategy shows a plateau of reasonable results, not a single sharp peak.
• Sub-periods: check results across decades and market regimes.
• Other markets: apply the same rules to other countries or asset classes where the mechanism should also apply.
• Cost stress: double the cost assumptions and see whether the strategy survives.
• Randomisation: run the strategy on shuffled or synthetic data to see what results arise by chance.

Stage 7: the hold-out test

Keep a final period of data that has not been used for any decision. Run the frozen strategy on it once. If the result is much worse than in-sample, resist the urge to adjust and retest on the same period, which would turn the hold-out into in-sample data. Record the result, revisit the hypothesis, and if necessary start again with a fresh hold-out. Statistical tools such as the deflated Sharpe ratio help judge whether in-sample performance was credible given the number of variations tried.

Stage 8: paper trading

Run the strategy in real time on simulated capital, generating signals from live data and recording hypothetical fills. Paper trading catches problems that historical tests miss: data that arrive late, corporate actions handled differently, signals that depend on revised data, and operational mistakes. It also provides genuinely out-of-sample results, though usually over too short a period to be statistically conclusive. Our guide to automated trading systems covers the infrastructure involved.

Stage 9: monitoring and retirement

Define in advance what would show that the strategy no longer works: a drawdown beyond a multiple of the historical maximum, a sustained drop in the signal's predictive power, or a change in market structure that removes the mechanism. Monitor these indicators routinely. Strategies decay as they become known and crowded; a written retirement rule prevents the natural tendency to hold on and hope.

The role of openness

Each stage produces artefacts that can be published: the written rules, data scripts, backtest code, factor regressions, robustness tables and the hold-out result. Publishing them invites the external review that catches errors an author cannot see. It also creates a public record of how many attempts were made, which is the most important context for any result. The organisational side is covered in running a research project in the open.

Openness does not remove the need for judgement: reviewers can be wrong, and a strategy that survives public scrutiny can still fail live. But it shifts the burden of proof onto the evidence rather than the author's confidence.