A two-gate held-out test of hand-labeled pullback nets

Two evaluation gates

Editorial framing: evaluate this workflow as two separate gates. The first gate treats visual pullback reading as a frozen labeling task. The second gate wraps the resulting score in fixed entry, target, and time-stop rules and scores that whole procedure only on dates that were never marked or tuned.

The archive workflow trained two networks as pattern recognizers, one on hand-marked pullbacks in uptrends and one on hand-marked pullbacks in downtrends. A pullback-in-trend was treated as a small counter-move inside an established directional move, and as a candidate re-entry rather than as a forecast of the next bar.

Gate one: a frozen binary-mark-target

Each bar received a binary-mark-target of 1 when a human marked that bar as a pullback example and 0 otherwise. That mark was the training target. It was not an order.

Network inputs were spaced-lookback-differences: close-to-close changes taken on a gradually lengthening lookback schedule, so recent action stayed dense while longer context remained present. Each difference was then turned into a range-normalized-input by dividing it by a 100-bar average true range, so the pattern inputs stayed comparable across volatility regimes.

The fitted network had 20 inputs, five middle-layer units, and one output. Training stopped when correlation on the internal-test-split began to decline. The reserved window was not searched.

What stayed in sample, and what stayed sealed

Facts through 1 June 1995 were used to train and internally check the networks. The interval from 1 June 1995 through 27 June 1997 was reserved as a sealed-date-window: unused until the finished mechanical procedure was evaluated once.

Inside the in-sample fact file, 80 percent of facts were assigned to training and 20 percent to an internal-test-split. That split was used only to stop training. It was not the reserved out-of-sample window.

Gate two: a mechanical-wrapper on the same inputs

A mechanical-wrapper turned the network score into one testable entry-and-exit procedure. The long-side procedure bought at market when network output exceeded 0.1, placed a limit exit at entry plus 3.5 times a 10-bar average true range after more than one bar in the trade, and forced an exit after more than 18 bars.

The short-side procedure reused the same labeling and input pipeline but inverted the orders and cut the maximum hold from 18 bars to 2 bars, because downward index moves were judged to be faster.

A single run-mode switch either wrote a preliminary fact file of inputs plus bar timestamps or, after training, fed the same inputs to a saved network and issued orders. Entry, exit, and abstention therefore ran as one mechanical procedure.

One unused-window score

Out-of-sample evaluation on the sealed-date-window counted 35 long-side trades and 18 short-side trades. Reported mean-profit significance was 0.0179 on the long side and 0.26 on the short side after one unused-window test each.

Editorial reading: those figures describe one historical score of the finished wrapper. They are not a claim about present-day use, and they are not a reason to retune the reserved window.