Walk-forward hybrid rules for intermarket forecast stacks

Forecasts as an information system

Forecast networks can be deployed as an information system that emits next-session price, direction, or turning-point estimates. The estimates may come from one network or from several networks working together.

Independently trained networks can estimate the next session's high, low, short-term trend, and medium-term trend. Those estimates can be read separately or used to confirm one another.

In a hierarchical forecast stack, the lower-level estimates can be passed as inputs into another network trained only on a further target, such as turning points. Each network then learns a single output.

Hybrid rules and stand-aside signals

A network trained to emit buy, sell, and stand-aside signals will reproduce the developer's chosen trade points, inputs, and preprocessing. Those signals need not fit a different capital base or drawdown limit.

A hybrid rule overlay can keep the network as a forecast source and let a separate rule set generate the actual signals. That rule set may run from simple formulas to a full expert system, so entry, exit, and abstention remain explicit. A stand-aside signal is then part of one testable procedure rather than an afterthought.

Intermarket input sets and other development choices

An example build covered yen, Treasury bonds, Eurodollar, and the S&P 500. For each target, one network set predicted the next-day change in the high from internal price, volume, and open-interest inputs. Another set added an intermarket input set.

Development choices included a feedforward back-propagation paradigm, a sigmoid transfer function, and a single hidden layer. For the yen intermarket set, inputs came from the Nikkei, Treasury bonds, Swiss franc, Deutschemark, US Dollar Index, Eurodollar, and British pound.

Preprocessing used price differences, simple and exponential moving averages, stochastic indicators, and intermarket spreads. Inputs were clipped beyond a set deviation threshold and then linearly scaled.

Next-day high estimates were described as information for stop placement and as a cue to next-session resistance. Complementary search methods such as genetic algorithms were noted for optimizing network parameters rather than replacing the whole procedure.

Walk-forward holdout as a system-optimization choice

A walk-forward holdout trains on an earlier fact window and evaluates reserved facts at fixed intervals during development. In the archive workflow, training held the learning rate fixed, omitted momentum, evaluated several error measures at set intervals, and retained the best network for each measure through an error-statistic checkpoint.

Architecture, inputs, preprocessing, and training settings are system-optimization choices. They belong in the same walk-forward procedure as the hybrid rule overlay so the full system, including stand-aside behavior, stays testable.