Constructing nearest-neighbor forecasts gated by a trend filter

How the forecast is formed

A nearest-neighbor forecast is built by locating historically similar patterns and averaging those neighbors’ subsequent moves to estimate the current pattern’s next move. Historical analog comparison ranks earlier windows of ordered observations by distance from the current window and uses the neighbors’ subsequent moves as the forecast raw material.

Treat the lookback as a searchable pattern

Pattern recognition treats a chosen lookback of prices or derived series as a multi-element signature that can be matched, classified, or assigned element-level importance. The ordered inputs that form the pattern may be past prices or other series such as oscillators and moving averages, so each pattern element can be assigned a distinct importance weight.

The lookback that defines the pattern window can be fixed by the builder or searched jointly with which series and parameters enter the pattern.

Rank analogs and form the next-move estimate

Neighbor contributions can be scaled so closer analogs receive more weight than distant ones. Distance-weighting applies that scaling, and the decay of weight with distance can be set separately for each pattern element.

Classify instead of estimating a next move

The same analog machinery can classify the current pattern into discrete act-or-stand-aside states instead of emitting only a continuous next-move estimate.

One weekly construction path

One construction path writes ordered weekly closes, optionally as the natural logarithm of the close, ranks analogs in a table after skipping the current bar, then marks the five nearest neighbors back on the price series.

Build the trend filter as its own module

A trend filter is a directional gate assembled from the signed and absolute close displacement over a chosen period. The construction compares the one-window and two-window sums of those displacements and holds the prior directional state when the index is not positive.

Compose the gated forecast

Analog ranking of a current price pattern and a period-based trend-detection status can be built as separate modules and then composed when constructing a gated forecast.