Stacking scored filters into a hierarchical stock outlook system

A rule-based outlook engine

The construction target is a rule-based expert system that issues medium-term and long-term US stock-market outlooks. Medium-term means three to six months. Long-term means six to 12 months.

Those outlooks come from a modifiable if-then-else knowledge base and an inference engine that combines the weighted conclusions.

How rules become facts

When an if-condition holds, the then-statement becomes a fact. When the if-condition fails, the else-statement becomes a fact. Else branches are optional.

Small trend filters with explicit weights

A price trend filter treats the NASDAQ Composite above a rising 200-day exponential moving average as a long-term bullish conclusion. The same conclusion applies when the 200-day average is now neutral but had been falling. That bullish conclusion has weight 20 on a 0-to-50 scale and sets a downstream variable K to 4.

A monetary rule labels the long-term outlook bearish with weight 20 when the three-month Treasury bill rate exceeds the long-bond yield times 1.03. Otherwise the rule records a normal yield curve.

A quiet-volume bull-market filter

The negative volume index is a volume-conditioned price series. It starts from a nominal base such as 100 and adds the index percent change only when volume falls versus the prior period.

The spreadsheet-computed trend is read by the expert system and scored on a plus 6 to 0 scale as a bull-market trend filter.

Hierarchy and backward-chaining

The preferred construction uses many simple hierarchical rules and backward-chaining. Only the lowest layer takes user or database input. Higher-level facts are derived from lower-level rule outputs.

Increment, net and normalize

Four competing outlooks are scored: medium-term bullish, medium-term bearish, long-term bullish and long-term bearish. Increment-decrement scoring adds or subtracts points for each fired outcome instead of averaging a small probability scale.

Bullish and bearish totals are then netted. For example, 230 bearish versus 50 bullish yields 180. Normalization converts that net outcome to a 0-to-100 range against the maximum possible weight. The normalized value is mapped to one of nine qualitative levels.

Composite indicators in a later rebuild

A later rebuild groups lowest-level indicators into monetary, sentiment and market-structure composite indicators. Each composite is scored on a symmetric scale from plus 6 to minus 6, then fed into higher-level forecast rules as a single fact.

Composite totals are mapped to a 0-to-100 range with Y=100*(X-A)/(B-A). Overlapping thresholds allow mixed membership before a forecast rule fires. For example, sentiment 58-70 is bullish and sentiment 48-62 is neutral.

Mechanical checks and historical reruns

The mechanical trading system is a closed if-then-else procedure that turns rule inputs and market state into one of a fixed set of medium-term or long-term outlooks.

After automatic checks for empty conclusions, missing facts and loops, the procedure is rerun on historical sentiment, monetary and market-structure series around major turns. Weekend exception updates take 30 to 40 minutes. Later work is specified as statistical indicator tests plus more historical forecast tests.