Coded rules should face one test, not a kinder sample

Write the procedure, then let it trade

A mechanical trading system is created by encoding the rules and then letting those coded rules place the trades. Mechanical design is contrasted with discretionary timing: building the system means writing code so the procedure, not a moment-by-moment judgment, fires each trade.

In that framing, a mechanical trading system is a complete set of coded entry, exit, and abstention rules that generate signals without a discretionary override at each decision.

Parameterization is part of the design

Parameterization is the choice of numeric inputs, such as moving-average lengths, that define how a rule set behaves across changing conditions. It is framed as a general design problem, illustrated by choosing faster and slower moving-average lengths and then changing those periods as market conditions change.

Walk-forward optimization is included among the validation topics presented with parameterization. That validation style re-estimates parameters on successive in-sample windows and then tests them on later unseen periods.

Inspect the system report

After a backtest, the system report is treated as a primary inspection surface. Accuracy and the ratio of average winning trade to average losing trade are singled out as two of its more important fields.

Short-hold systems lasting a few days are described as typically showing higher accuracy and a win/loss ratio near 1. Systems that hold a few weeks seek large winners to offset large losers. Around 60 percent accuracy with smaller profits is presented as a practical expectation that still requires sizing for drawdowns.

Shared ideas, including a parameter-free pattern

A bond procedure is built from 16 patterns that share three ideas: exit at the first profitable open, infer direction from bar-to-bar relationships, and filter with day-of-week seasonality. At least one pattern is written using no free parameters.

Hunting an optimal lookback

Tables of profit versus lookback period across a range of parameter values are used to argue that hunting an optimal lookback is a fallacy. Curve-fitting is the central risk of that search: selecting parameters that look profitable mainly because they were tuned to the same history used to judge them.

System optimization is the search of lookback windows and other parameters against historical results, including the risk that the chosen values only fit the sample.

Mean reversion as a premise

Mean-reversion systems are defined as those that assume a true value overshot when the market is overbought or oversold. The design premise is that price can overshoot a fair value and later revert after those extremes.

One parameter-free example uses a centerline from the square root of high times low, buying if the next open is below it and shorting if the next open is above it, with exits created by reversing the position.

Editorial view: that centerline is still part of the same procedure as the reverse exits. It should not be treated as a separate knob that can be fitted after the entries already look kind.

Drawdowns and stop-losses

Drawdowns are treated as a reason traders abandon a still-running procedure. A drawdown is an equity decline large enough that many traders leave before later trades can occur. A sizing rule of thumb is offered so an account can survive a worst-case decline.

A separate analysis section presents stop-losses as one of several ways a rule set can lose money. Editorial view: a stop is another coded rule in the same procedure, not a patch added after the sample has been made to look kind.