AI and Elliott search: a validation protocol
A toy neuron, chronological splits and candidate filtering with explicit limits.
Published: (UTC+3)
Русский · English · Deutsch · Español · Français · Italiano · 繁體中文 · 日本語 · 한국어
Automatic counting, structural search and a trained neural network are different mechanisms. The following invented inputs make their checks reproducible; they do not establish which code is used inside commercial product versions.
AI and neural networks: outputs and time splits
The toy neuron z=0.5 x−0.5, y=1/(1+exp(−z)) gives 0.622459 at x=2. This is the output of a chosen function, not a measured probability of wave completion. Its weights are assigned manually; no training, calibration or market accuracy is asserted.
For a two-bar target horizon split 100 bars into training 1–60 (signals 1–58), validation 61–80 (61–78) and test 81–100 (81–98). Features use available history only; fit normalization on training alone. Freeze version and parameters before testing, retain label availability times and include costs. A neural label does not remove future leakage from training annotations.
Candidate search: constraints before scores
Define the range, candidate cap and pivot availability first. Reject hard-constraint violations before ranking survivors. Keep unknown subdivisions unknown. A score does not override a rule or become a success probability.
In the invented six-point example, candidate A has wave 4 at 119 and score 0.9: reject overlap. B at 125 with score 0.6 passes the price subset only. Hitting a search cap does not prove no other structure exists. Save rejection reasons, version, inputs and cutoff; this published example is not claimed as Pro’s source algorithm.