Neural-net pairs from a prediction spread

A two-step pairs procedure

The case study frames pairs-trading as a two-step system. A neural net first predicts one stock from another name in the same sector. The gap between the actual price and that prediction then produces simultaneous long and short hedge signals on both names.

The method does not cross two normalized prices. It uses one stock to forecast another and times trades from the size and sign of the actual-minus-predicted difference. That difference is the actual-predicted-spread, and it becomes the entry, exit, and stop trigger.

How the predictor is chosen

A genetic optimizer can pick the predictor from a sector basket and later customize buy and sell thresholds for that pair. The case study does not rely on a fixed two-standard-deviation rule.

This is neural-network-pair-matching: a constrained one-input network lets the optimizer, not a prechosen pair, select which related security best predicts the target.

In the utility-sector example, the optimizer was limited to one input and zero hidden neurons. It therefore had to choose either Pepco or Southern Company as a linear predictor of Consolidated Edison.

After training, the network assigned Pepco a 100 percent contribution-factor and dropped Southern Company as the input for the Consolidated Edison prediction. With only one allowed input, that contribution-factor is the full share of the selected predictor in the trained model.

Why the forecast stays loose

The authors required a loose linear fit, not a highly accurate nonlinear one. The aim of loose-fit-prediction is that actual and predicted values still cross and then separate enough to define a tradable spread.

Because the actual close and the prediction sit in the same price range, the rules use the absolute difference of those two series. They do not normalize the two stock prices.

Shared rules for the hedge

Both legs share the same proxy-entry parameters so entries fire together. The implementation keeps a +1 or -1 side flag unoptimized to track the simultaneous hedge. The hedge is therefore entered and exited as one pairs-trading procedure.

Search ranges supplied for the shared rules were an entry spread of 2 to 4, an exit spread of 0.01 to 1.9 as the series reconverge, and a rarely hit stop exit of 4.1 to 6.