Constructing an intermarket neural net trading system

The historical workflow built a Mechanical trading system in which a network was only one stage. Daily date, open, high, low, and close futures histories were used for training, while a shortest-path interpolation from open to close through that day’s high and low created simulated ticks used only for later hypothetical buy and sell tests.

Intermarket input search

When a network was trained for one futures market, every other series in the multi-group database was offered as a candidate input, both inside and outside that market’s group, and training kept only the series judged useful.

Locked data transforms

Raw prices were not fed in unchanged. Change was expressed as a logarithm of the move between two times, location versus a recent mean was expressed in standard-deviation units, and both transforms were then scaled into the continuous interval from -1 to +1.

Range forecasts, not a binary call

The networks were trained to emit point estimates of highs and lows over several horizons so that a range of up and down move sizes, not only a binary direction, could be turned into probabilities, targets, and stops.

Forecast skill was judged from at least 30 recent prediction-versus-outcome pairs, with average error expected near zero. The standard deviation of those errors, not their absolute size, was used to place the point estimate on a normal curve, with about 68 percent of outcomes expected inside one standard deviation and about 95 percent inside two.

A mechanical payoff table

Entry, target, and stop choices were framed as a mechanical payoff table. Actions were ranked by long-run expected economic consequence given event probabilities rather than by indicator crossovers, and a very close stop was treated as likely to be hit and therefore a near-certain loss.

After expected return was computed from target, stop, commission, and stop slippage, candidates with negative expectation or unacceptable ruin risk were dropped. Remaining trades were scored by expected return times the chance the entry would print, then the best score across look-ahead horizons was chosen by enumerating a large decision tree.

Editorial: that ranking and horizon search is the System optimization step. The network supplies probabilities; the Mechanical trading system decides whether an order is allowed.

A held-out significance gate

After training and rule definition, the assembled procedure was simulated on historical prices not shown during training, under a pessimistic null that any positive sample mean was luck until a t-test against a zero-mean random-selection distribution rejected that null at a 95 percent confidence level, with a critical t near 1.645 when the simulation had 30 or more trades.

Editorial: that check is the Walk-forward analysis gate. No signal is allowed to fire until the assembled procedure, not the network alone, clears it.

What construction did not promise

Building the software and trading procedure was described as requiring substantial research, design, programming, and testing. No method, including this one, was expected to produce near-certain winning trades, and relationships used in past rules can stop working when market dynamics change.