Skip to main content
Track System optimization
12 / 23
Search the archiveLibrary
Library

1997issue C011-9

Rule-based intraday system design with genetic search

A mechanical trading system can be written as two thresholded price-comparison rules plus a genetic search that tunes their inputs together. Solutions are judged by whether entries, exits, and abstention stay consistent on held-out periods, not by in-sample profit alone.

  • The mechanical trading system enters the S&P 500 on the next open and exits at that same session close only when two thresholded price-comparison rules are both true.
  • System optimization searches six inputs together with recombination, mutation, and selection instead of stepping every parameter combination.
  • In-sample fitness is net profit on the training window, but solution choice also checks whether profitable trades are spread through the test period and whether withheld spans stay consistent.
  • The reported systems take only intraday trades, are framed as multi-system portfolio candidates, and point toward letting the genetic search choose the rules themselves.
Entries in this reading3 entries
Current trackSystem optimization131 readingsRule-based entry292 readingsMechanical trading system322 readings

A mechanical trading system as one procedure

The mechanical trading system is an open-to-close S&P 500 procedure. It enters on the next open and exits at that same session close when two thresholded price-comparison conditions are both true. If either rule fails, the system abstains for that session.

Each rule compares two selected prices from the recent past and requires the first to exceed the second by a multiple of the average absolute difference between those prices. Rule-based entry, exit, and abstention are therefore tested as one object rather than as separate decisions.

Six inputs searched together

Six parameters are searched together: for each of the two rules, two lookback indices that pick the prices and one threshold multiplier. A genetic search that uses recombination, mutation, and selection is used instead of stepping every parameter combination, because a full grid over six inputs would produce an enormous number of trials.

The six inputs are therefore treated as system optimization of one mechanical trading system.

How solutions are chosen

In-sample fitness is defined as net profit on the training window. Solution choice also inspects whether profitable trades are spread through the test period rather than bunched in a few blocks.

The historical series is continuous back-adjusted end-of-day data, split so one window is in-sample and two other spans are withheld for out-of-sample checks. The author prefers solutions with fewer trades or less profit per trade when they show lower drawdowns and more consistent trading than the optimizer's raw fitness ranking.

Comparison with a hypothesis-based model

Relative to an earlier hypothesis-based model that allowed very tight stops but thin profit per trade, this empirically tuned model is described as having adequate profit per trade while being unable to use tight stops and showing higher drawdown.

The reported systems take only intraday trades and are framed as candidates for a multi-system portfolio rather than a finished standalone method. The planned next step is to let the genetic search choose the rules themselves, not only the parameters of two fixed price-comparison rules.

Half-year net profit of two genetically searched long-only S&P 500 rules

The in-sample fitness champion posts large 1988–94 gains and then gives them back in the 1996 holdout, while the best-overall long rule stays profitable in most half-years on both sides of the sample. Net dollars for each six-month bucket are taken from the article’s printed long-side performance summaries.
The in-sample fitness champion posts large 1988–94 gains and then gives them back in the 1996 holdout, while the best-overall long rule stays profitable in most half-years on both sides of the sample. Net dollars for each six-month bucket are taken from the article’s printed long-side performance summaries.S&P 500 futures · Half-year · 1986-01-01T00:00:00.000Z to 1996-10-31T00:00:00.000Z

Fitness was in-sample net profit from 4 January 1988 through 30 December 1994. Half-years in 1986–87 and 1995–96 were held out. Assumed costs were a $30 round-turn commission and $75 slippage.

Educational research material, not investment advice. Historical source context does not establish present-day performance.
12 of 23 in the System optimization track
19971-10 pp.Next on System optimizationEvolving three-rule systems with genetic templatesA complete mechanical system can be encoded so a set of numbers maps one-to-one onto fully specified rules, using the first scaled value as a template index and the remaining values to fill that template's blanks.
All readings on this track · 23 readings
  1. 1984Evaluating mechanical systems beyond net profit
  2. 1987Mechanical systems need more than pretty charts
  3. 1987Evaluating silver moving averages and RSI by scoring rule
  4. 1988Simple systems over optimized complexity
  5. 1988Cycle rule inputs until the parameter search stops
  6. 1990Split-sample tests for stochastic parameter grids
  7. 1993Four-subsystem trading system construction
  8. 1993Randomized price paths for stock system parameters
  9. 1994Mechanical timing with out-of-sample and filter checks
  10. 1995Evaluating asymmetric percent-reversal trading rules
  11. 1996Evaluating a daily volume-strength and new-high new-low timing procedure
  12. 1997Rule-based intraday system design with genetic search
  13. 1997Evolving three-rule systems with genetic templates
  14. 1997Audit the holdout and the exposure before the fitted entries
  15. 2001Inner weights are part of the candlestick signal
  16. 2003Mechanical short-term rules for news-sensitive rallies
  17. 2004Band construction as a testable trading procedure
  18. 2006Layered day-session automation from brackets to rules
  19. 2006Why mechanical rules must be retuned per security
  20. 2006A linear-regression next-bar forecast needs a searchable cutoff
  21. 2015Construct if-then rules, then backtest and optimize
  22. 2015Mechanical system design as a critique of discretion
  23. 2025Joint channel design with a trend filter and stop-loss orders
By fieldSystem design203 readingsBy year199791 readings
All 131 readings tagged System optimization
Also on System optimization5 readings
1995Evaluating asymmetric percent-reversal trading rules1997Evolving three-rule systems with genetic templates2003Mechanical short-term rules for news-sensitive rallies2006Layered day-session automation from brackets to rules1992Building TRIX crossover and momentum entries on a period grid