Request access to the classic-patterns screener models
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These models are shared for research and education only. They are not investment or financial advice and not a signal service; there is no warranty of any kind. Requests are reviewed by hand.
Log in or Sign Up to review the conditions and access this model content.
classic-patterns screener models
Shows pattern setups with estimated probabilities; you decide. Do not take every pattern it finds — use your own judgement. This is a research / education project and not investment or financial advice. No warranty.
Code, data download and the screener app: https://github.com/ai-arie/classic-patterns
What the models do
For a chart pattern detected on daily US stock data (cup-and-handle, deep cup, head-and-shoulders top, inverse head-and-shoulders, consolidation breakout, ascending triangle) that triggered at today's close, the models estimate how the pattern's own trade plan would play out: entry at the next open, the pattern's stop and target, at most 60 trading days. Output: P(target first), P(stop first), P(neither within 60 days) and an expected R (multiple of the initial risk, net of 10 bps per side, from training-period averages).
| file | model |
|---|---|
student.int8.onnx |
distilled student: Chronos-2-Small (autogluon/chronos-2-small, 28M parameters) fine-tuned with custom heads, distilled from three Chronos-2 teachers; ONNX int8, CPU, ~78 ms per stock |
student.onnx |
the same student, fp32 |
lgbm.txt |
LightGBM on hand-crafted pattern, stock and market features |
Both are calibrated (temperature scaling) on the last ~2 years of training data. Also included:
student.json / lgbm.json (inputs, calibration, settings), configs/ (the detector, feature
and label settings the models were trained with), manifest.json (code and data hashes),
parity.parquet (reference scores to check your installation) and checksums.sha256.
The screener app built on these models shows long setups only: short setups are in the training data but are not shown because their results in testing were not good enough. Setups that are still forming show only their chance of triggering.
How to use
- Install the screener from the code repository (
uv sync --extra data --extra app) and build your data as its README describes. - Once your access request is approved, log in (
hf auth login) and download this repository's files intomodels/production/of the code folder:hf download ai-arie/classic-patterns-models --local-dir models/production. uv run python scripts/parity_check.pymust printPARITY OK(your data build reproduces the reference scores), thenuv run python scripts/screen.py --refreshanduv run streamlit run src/classic_patterns/app/ui.py.
No PyTorch or GPU is needed.
Training data
- Daily prices of US stocks listed on 2026-10-23 (Nasdaq Trader symbol directory),
from Yahoo Finance via
yfinance; training samples from 1995-01-01, data to 2026-10-23. Training split until 2024-10-01, calibration / early stopping after. - Setups only on liquid days (close ≥ $5, 50-day dollar volume ≥ $15M, ≥ 300 days of history).
- No price data are included in the weights or in
parity.parquet, which holds 300 scored setups (ticker, date, scale-free geometry, scores) for the installation check.
Limitations
- Survivorship bias: only companies listed today; delisted stocks are missing.
- Yahoo data quality (adjustments, gaps) is not guaranteed.
- Costs are a flat 10 bps per side; short-borrow cost and availability are ignored.
- Pattern definitions are one operationalisation of subjective chart patterns.
- At screening time the entry price is unknown; the models use today's close in its place.
- Targets are the patterns' textbook measured moves and are usually far away. Setting take-profit levels was not trained, so use your own judgement, especially with the targets.
- Probabilities describe past behaviour of similar setups; markets change.
- Chronos-2's pretraining data are not completely documented.
Compute
The whole project (hyperparameter search, walk-forward evaluation, holdout test, final models) used approximately 141 GPU-hours on NVIDIA A100 80GB GPUs; the final models themselves about 5 GPU-hours. The app runs on CPU.
Intended use
Research and education: screening and studying chart patterns. Not for automated trading, not a signal service, not investment advice.
License
Apache License 2.0 (see LICENSE and NOTICE). The student derives from Chronos-2-Small
(Apache-2.0, Copyright Amazon.com, Inc. or its affiliates).
Model tree for ai-arie/classic-patterns-models
Base model
autogluon/chronos-2-small