Cost-aware walk-forward evaluation of pattern-detector signals

Define the mechanical trading system

A mechanical trading system is one testable procedure that turns rule inputs, market state, and execution constraints into entry, exit, or abstention over the system holding period.

A limited-connection network scores price momentum, price acceleration, volume movement, and price volatility in separate first-hidden-layer channels, then a fully connected second hidden layer combines those detections into one output.

Price velocity is the change from the price five days earlier, and price acceleration is the five-day difference of that velocity. Each series is presented to its own input group.

The preprocessor was designed around 300 to 400 weeks of relevant history. Limited tests on series that lacked volume were used to argue that volume-movement patterns should remain in the input set.

Lock the training label and the zero band

The training label is an exponential moving average of five-week relative gain, with average price defined as the midpoint of the weekly high and low. The five-week gain target is that smoothed relative change from current average price to average price five weeks ahead.

Inputs are smoothed and then, with the training target, rescaled into the range from -0.5 to +0.5 so the network is more sensitive near zero, where the buy and sell rules trigger.

Smoothing is used to cut noisy forecasts and false action signals, but it also delays those signals. Delayed crossings can themselves be false, so input parameters are tuned against that trade-off.

Update weights without inviting paralysis

Weight updates follow Hebbian strengthening plus back-propagation of forecast-versus-target error. The stated safeguard against fitting the sample is at least four training cases for each weight.

Simulated annealing is used to start weights near a low-error region so later back-propagation needs fewer iterations and is less likely to enter network paralysis. Network paralysis is a training state in which large competing weights keep cutting in-sample error without improving later forecasts.

Walk the forecast forward, then charge the pair

The evaluation protocol trains, forecasts the five-week gain, then advances one week and repeats. The walk-forward step lets forecast and target be compared over time. The recorded forecast can lag the smoothed target and can emit a reversal before the later price path confirms it.

A rule-based entry is a signal that fires when a processed forecast or training label crosses through zero, rather than at an exact price extreme.

Completed long pairs were charged a combined slippage and commission rate of 1 percent. Short sales were omitted, and a stop-loss overlay was not applied because it would reshape the outcome histogram. Commission analysis is an evaluation filter that charges implementation cost to every completed pair before a signal is treated as executable.