Constructing five-session forecasts from stochastic, ADX, and MACD inputs

Specify the output before any weights move

A market predictor of this type is stored as layers of units whose numeric connection strengths are adjusted during training until the architecture maps examples to a specified target.

Whether the trained system is usable depends on how the builder defines the input architecture and the desired output, because the finished weight matrix does not expose an inspectable reason for each decision.

Editorial reading: treat that dependence as a specification problem. Lock complementary lookbacks, an MACD histogram definition, and a five-session target before any network is allowed to fit weights.

Lock complementary lookbacks and a five-session target

The illustrated input block includes a nine-session slow stochastic %K and %D pair and an 18-session average directional index. The stochastic oscillator is a slow %K and %D pair computed over a nine-session lookback and supplied as ordered oscillator inputs to a forecast network. The average directional index is an 18-session directional-movement reading used as a trend-strength input rather than as a standalone chart rule.

An MACD histogram input is formed as the difference between a 12-versus-26-session exponential MACD line and a nine-session signal of that line. That histogram is used as a chart-derived input.

Two further inputs are the current index close and the close five sessions earlier. The designated target is the percentage change in that index five sessions ahead. That forecast horizon is a fixed five-session lead that defines the network output as a percentage change, not as a contemporaneous classification.

Treat near-duplicate series as a design risk

Financial inputs that are slight variants of one another can hinder training. Stacking more series of the same type is treated as a design risk.

Input redundancy is overlap among indicators that recode nearly the same price path and can impair training if they are stacked without a construction reason.

Add a sized hidden layer, then run the training phase

One construction choice replaces a three-layer default with four layers and sizes the added hidden layer at about 12 percent of the input-layer unit count. The hidden layer is an intermediate layer of weighted units that transforms the indicator block before it reaches the forecast output.

The training phase is the interval when connection strengths are raised or lowered according to whether a guess matches the designated target.

After most training examples produce the intended output, the frozen weights are used as a predictor. The same net can also surface indicator combinations that a later rule system might encode.