Randomized price paths for stock system parameters

Independent inputs for setup and trigger

A mechanical procedure is specified as two parts: a longer-horizon setup and a shorter-horizon trigger. Those parts must use independent raw inputs so the same price change is not counted twice.

The setup is the longer-horizon condition meant to say whether the market is positioned to rise or fall. The trigger is the shorter-horizon rule that times entry once a setup is present and should respond to recent price change.

Setup ingredients and position size

Longer-horizon setup inputs are grouped as interest-rate policy, sentiment extremes, commercial versus public open-interest activity, and broad-market momentum sampled on a longer interval than the trigger.

Interest-rate policy shifts and sentiment extremes are described as too infrequent to serve alone as an active setup. Weekly market momentum combined with commercial-versus-public open-interest readings is presented as producing more usable setup events.

Position size is described as scaling with the direction and intensity of the longer-horizon indicators, including cases where policy, sentiment, and setup signals align.

Forward-testing and short samples

Rule design is given priority over hunting for optimum lookbacks. Forward-testing means choosing parameter values on one historical segment and evaluating the same rules on a later unused segment. Ordinary forward-testing is presented as needing a long sample, given as 15 to 20 years for intermediate-term systems, few fitted parameters, and segments large enough to include cyclical and secular variation.

When many parameters must be chosen from a short history, a seven-parameter example on a 10-year sample is given. The procedure randomizes that history, estimates values on the synthetic series, and repeats the cycle so the chosen values reflect variation around the original sample.

Rebuilt bars and an rsi-adaptation

An rsi-adaptation is specified with three operating parameters: lookback length, comparison-close lag, and distance from the 50% midline, plus a money-management stop. It is an oscillator derived from the relative strength index that adds a comparison-close lag and a band around the midline, and it is run as a reversal system in which a new long or short signal ends the opposite position unless the stop is hit first.

A randomized-database is a synthetic bar history rebuilt by resampling range, open and close location, and low-to-low change from the original series. Each synthetic daily bar is rebuilt by independently resampling, from a lookback of 1 to 30 days, four statistics of the original series: open location in the day's range, high-low width, low-to-low change, and close location in the day's range.

A sameness-check keeps a synthetic series only when a t-test finds the real-versus-synthetic difference indistinguishable from zero.

Grouped search on randomized databases

A comparison design fits parameters on 10 to 20 randomized sets, and separately on each real half-sample, and then evaluates those values on the unused real half. Search progress for grouped parameters is plotted against an average score computed over 30 randomized databases, reverting to the last improving group when a trial is worse.

For systems with many parameters, the search is partitioned into independently tradable groups for the trigger, the setup, and money management, plus related subsets. A parameter-group is a logically independent subset of rules that can be searched while other subsets stay fixed.

Nominal values are first obtained group-by-group on the real sample. Only one group is then allowed to vary while the full system is scored on randomized databases, with longer-horizon indicators assigned the sizing role.