Constructing walk-forward neural trading rules

The network as one procedure

A back-propagation unit forms a weighted sum of incoming values, adds an adjustable threshold, and applies a saturating transfer function such as a sigmoid. Hidden units sit between environmental inputs and emitted outputs. A fully connected layout feeds every unit from every unit in the prior layer.

The construction treats this stack as a single mechanical procedure whose later stance is long, short, or stand aside, not as a discretionary overlay added afterward.

Fitting, reserved cases, and redesign

During fitting, a fraction of the gap between emitted values and designer-set targets is sent backward to update weights. Reserved input-target pairs are later presented with those updates turned off.

If reserved-set results are unsatisfactory, redesign options include incremental refitting, a full restart, revising which inputs or targets are used, and changing how many units sit in each layer.

The mechanical-signal and the walk-forward-split

One mechanical mapping treats two outputs as long and short indicators. It takes a long stance when the long indicator exceeds 0.5 and a short stance when the short indicator exceeds 0.5. That mechanical-signal is a predetermined mapping from network outputs to long, short, or no position, applied the same way while fitting, on reserved cases, and on later live inputs.

A walk-forward-split is the cutoff that ends weight updates. Later cases are scored with frozen weights and no error sent backward. In one weekly-index construction, after that cutoff the charted lines switch from training targets to the frozen network outputs.

Fitting targets may include later price action that the live procedure cannot see. A future-knowledge-target is that kind of fitting goal. After weight updates stop, later signals depend only on whether historical regularities were captured.

Input-normalization and precomputed series

Unscaled price and volume ranges that sit far outside a unit interval can pin a sigmoid. Construction practice is to remove level bias and rescale inputs toward 0 to 1 or -1 to 1. Input-normalization is that recentering and rescaling, so rule inputs occupy the sensitive region of a saturating transfer function instead of pinning it.

Supplying precomputed series such as moving averages, RSI, percent R, stochastics, or MACD as inputs is proposed so hidden units spend capacity on broader regularities instead of rebuilding those calculations.

Overtraining and capacity-revision

Overtraining is described as learning the fitting cases too exactly. It is the adaptive analogue of overfitting a conventional rule set: the fit reproduces the training cases too exactly and then fails to transfer to reserved cases. Suggested countermeasures include more varied cases, slight noise on successive presentations, and not inflating hidden-layer size without regard to memory and run time.

When reserved-set results fail, changing how many units sit in a layer is a capacity-revision. It is redesign of the same procedure rather than a new discretionary overlay.