Neural-net inputs and rule trees for mechanical systems

Write the trade as one procedure

Construction is described as marrying market-analysis judgment with the training algorithms. Models built by computer scientists alone, or by traders who do not understand the training step, both fail.

The first product is a mechanical-trading-system: a complete, testable procedure that turns rule inputs, market state and execution constraints into entry, exit or abstention signals over the system's holding period.

The intended construction loop is to encode the same entry, exit and abstention decisions a trader would make, then ask the network to reproduce that procedure rather than dumping every available series into the model.

Expert rules and neural networks

Expert systems are described as rule-driven decision trees whose path and combined certainty change with new facts, while remaining non-adaptive because the rules themselves are predefined. An expert-system activates rules by certainty factors and can change its path when new facts arrive, but does not rewrite its own rules.

Neural networks are described as empty until demonstrated examples are shown, after which they can find patterns and reweight old versus new relationships when new information is presented. A neural-network is a pattern-finding procedure that starts empty, learns from demonstrated examples, and can reweight old versus new cause-and-effect relationships when new information is shown.

A mechanical system is framed as choosing the method that fits the task, whether expert rules, a neural net, or even linear regression, rather than treating any one technology as a cure-all.

What linear tools miss

Linear tools such as regression are said to miss changing, nonlinear cause-and-effect among related markets that can be used to measure influence on equities. That missed structure is nonlinear-causality: inter-market cause-and-effect that is not a straight-line relationship and can change over time.

Sieve the inputs before the signal

Large candidate data sets, described as more than a million pieces of data per security, are not used wholesale. Techniques are applied to decide which information to keep, following a keep-it-simple rule.

Neural-network construction is described as better when kept simple: start with many series, sieve them down, and avoid building a network with on the order of 200 inputs. That habit is the input-sieve, the construction step that reduces many candidate data series to a small set before the network or rule tree is allowed to produce a signal.

System-optimization is the matching selection step. It keeps only those inputs and rule combinations that survive tests one at a time and then together, so the procedure stays simple enough to remain testable.

One described market-level expert system applied about 267 rules of market behavior and performed about 6 million calculations per decision, while more than 1,000 securities were analyzed alongside the market. The construction rule is still not to feed every series into the decision.

Admit a regime cue only if it changes the signal

Put-call ratio, sentiment, price inertia or momentum, and fundamental or macroeconomic series are listed as candidate inputs that should be written into a trading plan and tested in the network one at a time, then in combination.

Put-call-ratio is a sentiment input used as a market-regime cue. It can be admitted to the procedure only if it changes the signal when tested alone and in combination.