How a systematic strategy is built

A strategy is a complete decision process

A systematic trading strategy is a set of rules that turns information available at a particular moment into a position, then explains how that position will be managed and closed. The word systematic does not mean the strategy must be complicated, automated, or mathematical. It means that two researchers given the same data and the same rules should reach the same decision.

That is a much larger object than an indicator. A moving average, relative strength index (RSI) reading, volume-weighted average price (VWAP) distance, or order-flow imbalance is a measurement. It can describe something about the market, but it does not decide whether to trade. A strategy must still say which instrument is eligible, when the measurement is read, what other conditions matter, which side may be entered, how large the position is, what ends the trade, and which costs and fill assumptions apply.

This distinction matters because vague ideas acquire accidental rules during a backtest. “Buy when momentum is strong” can quietly become “buy this asset at the same close that confirmed a hand-picked lookback, hold until the result looks best, and ignore fees.” The computer will calculate that rule precisely, but precision does not make the research question coherent.

The parts that have to be declared

Most strategies can be understood as the following connected decisions:

The pieces are related but should not be blurred together. A volatility filter does not size the position unless the rule says it does. A stop does not define the original thesis; it defines one way of limiting or ending exposure. A backtest setting is not “just plumbing” when changing it changes which price could have been obtained.

A simple strategy, fully stated

Consider a daily moving-average continuation idea. The observation is that a market trading above its recent average may remain directionally persistent. The signal is the difference between a 20-day and a 100-day moving average. The strategy is the following larger sequence:

1. Use only instruments that passed the liquidity rule before the test period.
2. Calculate both averages from completed daily closes.
3. After the daily bar closes, mark the instrument long-eligible when the 20-day average is above the 100-day average.
4. Submit the modeled entry no earlier than the next eligible market event.
5. Size the position from a declared risk budget using volatility known at the decision time, subject to a maximum exposure.
6. Remain long until the fast average crosses below the slow average or a separately declared protective rule closes the position.
7. Charge the same fees, spread, and slippage assumptions to every variant.

Nothing in this example says the strategy is profitable. Its value is that the researcher can now identify what is being tested. A faster lookback changes the signal. Same-close execution changes timing. Volatility scaling changes the position. A stop changes the exit distribution. Testing all four at once would make it difficult to know why the result changed.

Signal, state, and action are different

Three labels prevent a great deal of confusion:

• A signal is an observable calculation, such as a moving-average spread.
• A state is an interpretation used to make the strategy eligible, such as “the spread is positive and volatility is below its ceiling.”
• An action is the change in target exposure, such as moving from flat to long at the next eligible event.

A signal may remain positive for fifty bars while the strategy performs only one entry. Another strategy might rebalance every bar in proportion to the same signal. They use the same measurement but implement different strategies. This is why reporting only an indicator name is never enough to reproduce a result.

Strategy, study, run, and result are different objects

Four objects are easy to blur together:

This separation makes conclusions traceable. “The strategy improved” is too vague if the only change was a more favorable test period. “The study found a higher net return for the slower exit on the held-out window” identifies the rule, comparison, metric, and evidence boundary.

A strategy can exist before it has a backtest. A backtest can execute a strategy without establishing that it will work later. A result can be calculated correctly while answering a poorly designed study. Keeping the objects separate prevents confidence in the software from becoming confidence in the trading idea.

Build one decision at a time

A useful first experiment isolates the parent idea. Start with one transparent signal, one entry rule, one exit rule, simple sizing, and a credible cost model. That version is the baseline. Add a filter, stop, confirmation signal, or adaptive parameter only after deciding what improvement it is supposed to produce.

For each addition, ask:

• Does it change the hypothesis or only the implementation?
• Which observations does it remove or delay?
• Does it improve the full distribution, or only one headline metric?
• Is the benefit visible across a reasonable parameter neighborhood?
• Does it remain after costs and on later data?
• What result would persuade us to remove the extra rule?

This approach makes failure informative. If the simple rule fails, the parent idea may be weak. If the simple rule survives but a filter does not help, the filter can be removed without discarding the entire strategy.

What a backtest of the strategy can establish

A backtest can show how the declared process would have behaved under its data and execution assumptions. It can reveal trade frequency, path dependence, exposure, drawdowns, turnover, sensitivity, and which rule produced each action. It can compare variants on a common basis.

It cannot establish that the behavior will persist, that the modeled fill was available, or that the rules were chosen independently of the same history. Those questions require later data, stronger execution evidence, and honest accounting for how many alternatives were considered.

Common construction mistakes

• Treating an indicator threshold as a complete strategy.
• Allowing the entry to use a value that was not final at decision time.
• Changing entry, exit, sizing, and costs together, then attributing the result to one of them.
• Choosing risk limits only after seeing the worst historical loss.
• Failing to state what happens when long and short conditions occur together.
• Comparing a complex net strategy with a simpler gross baseline.
• Saving only the final parameters and losing the strategy version, dataset, and assumptions that produced them.

Further reading

• Brock, Lakonishok, and LeBaron, “Simple Technical Trading Rules and the Stochastic Properties of Stock Returns” (1992) — A canonical example of defining moving-average and trading-range rules precisely and testing them against explicit bootstrap null models. It is useful for rule declaration, but is not a complete modern multiple-testing protocol.
• White, “A Reality Check for Data Snooping” (2000) — Shows how selecting the best result from a family examined on the same history changes the inference, and develops a benchmark-based test for that search.
• Perold, “The Implementation Shortfall: Paper Versus Reality” (1988) — Separates the paper portfolio from the implemented result and places execution cost, delay, and missed trades inside the performance process.