Construct a decision procedure that revises itself

Why the first rule set is not the system

A usable decision procedure is constructed so it can revise itself when market conditions change, rather than stay fixed after the first rule set is written. Construction here means a scoring loop that can be updated from later outcomes, not a finished list of rules.

Match the variety of what moves prices

The construction goal is to give the procedure enough input variety to match the many factors that move prices, instead of compressing the problem onto a few technical readings. Narrow technical or earnings-and-dividend screens are treated as information-destroying shortcuts that leave the decision system underpowered.

Price history alone is treated as insufficient for direction. A richer pattern that also includes fundamentals, psychology, and news is the proposed basis for anticipating trend changes.

Recognize each policy before an action is taken

Each well-defined policy is treated as a distinct information pattern that the procedure must recognize before an action is selected. Those policies are buy, sell, hold, do nothing, and sell short.

Inputs such as market indicators, earnings trends, and psychological measurements are encoded as a feature vector, then scored against a weight vector for every policy. The largest discriminant is the action taken.

What the computer holds and what the person keeps

The computer is assigned the dense information patterns a person cannot hold, while the person keeps judgment, supervision, and the final use of the scored alternatives.

The same scoring core is chosen because it can take many features, still work when some inputs are noisy or missing, and store only the current weights plus the latest feature vector.

Revise the same weights from later outcomes

Because the process that generates price changes is poorly understood, initial weights are expected to produce many poor decisions until feedback from later outcomes revises them.

A perceptron-style update raises weights on features that improved results and lowers weights on features that worsened them, using a review of whether each carried-out policy was correct.

Keep construction, psychology, and pattern work in one loop

Editorial: that review is how a Pre-trade checklist and the Trading psychology process stay attached to Pattern recognition. The checklist names the action, including doing nothing. The psychology process is the habit of scoring and then revising. Pattern recognition is the comparison of the current feature vector with the weights. None of the three is treated here as a separate system.