Constructing a dual-net day-ahead index direction forecast

A forecast built from ordered stimuli

A pattern-recognition forecast is a quantitative model that maps a fixed lookback of ordered market observations to a forecast label rather than a discretionary rule. The map is built by presenting those ordered observations as stimuli, adjusting internal connection strengths until the mapped response is acceptably close to a chosen target, and then applying that trained map to unseen inputs.

Editorial: treat the day-ahead index forecast as a construction problem. Specify the detrend, train two related pattern recognizers on different series features, require agreement before a signal exists, and keep a later unused window so the built baseline can be compared with a true holdout.

Detrending before the map sees the series

Raw index levels are a poor representation for this construction. The series is first rewritten as a normalized change so the model compares relative moves rather than exact prices.

Detrending subtracts a moving average from the raw series so the model sees relative deviations instead of absolute price levels. The working preprocess subtracts a 10-day moving average from each close or high and feeds a rolling 10-day window of those detrended daily values.

Two features, one confirmation rule

Two models share that window but not the same feature. One is trained on detrended daily closes and the other on detrended daily highs. Dual-net confirmation issues a signal only when both independently trained models agree on the next interval’s direction.

The target is next-session direction, encoded as a long-the-open/flat-the-close label versus a short-the-open/flat-the-close label, not a point estimate of tomorrow’s price. The stimulus-response map is the learned association between that preprocessed input window and the desired output label.

How the two nets are trained

Training uses two years of historical observations. Each net makes a single pass over that training history. The close model uses an 85 percent learning rate and the high model uses a 60 percent learning rate. Learning rate is a training control that limits how completely a new association is written on first exposure so the model does not merely memorize noise.

Both nets use a 100 percent memory profile so older associations are not aged out. Memory profile is a retention setting that decides whether older encoded associations keep equal influence on later responses.

Both nets are specified with 300 synaptic statistics, a conjugate option intended for negatively sloped functions, no sigmoid postprocessing on the output, unit-coefficient linear conversion on the response, and unit-coefficient sigmoid conversion on unbounded stimulus data. Complex-valued encoding is used so many stimulus-response patterns occupy the same stored footprint as a single pattern.

A forward window after training

The following five months are reserved as a forward window to inspect whether the mapping generalizes. A forward window is a later sampling interval held out after training and used only to inspect generalization.

A model trained only on the associations it has been shown cannot be assumed to respond correctly if later market conditions introduce events absent from that input history.