Evaluating encoded candlestick sequence hypotheses

A candlestick is a compressed code

An encoded candlestick is a short label that replaces one bar's body and two shadows with size bins instead of a named pattern. In the archive workflow, each weekly bar was reduced to a three-letter body-shadow code that separately described its real body, its upper shadow, and its lower shadow.

Those three parts were placed into five standard-deviation bins running from two steps above a typical reading to two steps below. A standard-deviation bin is a discrete size class that places each candle part on a scale from unusually large to unusually small.

The candle-to-code step was already described as lossy compression, the information discarded when OHLC geometry is reduced to a short code. Extra statistical reduction can discard still more of the original bar structure.

A single-bar hypothesis is not a next-week forecast

Isolated weekly S&P 500 candles covering the years from 1962 to 2014 were judged to have no useful value for forecasting the following week's index move. Next-week percentage changes were judged weak as forecasts because their scatter was usually wider than the gap between the average change and zero.

Single candles were described as carrying almost no next-bar information. A single-bar hypothesis is the claim that one isolated candle, by itself, changes the odds of the next bar. The archive did not treat that claim as a usable forecast.

Sequence length and the sample-size conflict

Pairs or longer runs were the suggested unit for a testable hypothesis. A sequence hypothesis is the claim that two or more consecutive encoded candles form a repeatable condition worth testing. Two-to-four-bar sequences make codes distinct enough to test.

That same lengthening demands tens or hundreds of thousands of sequence observations, so uniqueness and sample size compete. That tension is the sample-size conflict: longer sequences create more distinct codes, yet those codes need huge archives before they appear often enough.

A weekly sample of 2751 observations was treated as too small to treat rare encoded signatures as statistically settled. The main exceptions to the weak next-week pattern were codes that appeared only once or only a few times, including several two-occurrence codes and one five-occurrence code.

Out-of-sample decay is part of the test

Ten years of daily bars across S&P names and two years of hourly bars, both scanned as three-candle sequences, were reported to leave some later-window results about half as large as the discovery-window results. Out-of-sample decay is how much a sequence's discovery-window result shrinks when the same rule is retested on later bars.

Editorial reading: an encoded sequence is only a candidate signal after you specify a sequence length, a sample-size floor, and an out-of-sample decay check. Until those three choices are written down, the body-shadow code remains a compressed label, not a finished rule.