Candlestick pattern detector (YOLO11n)

Detects 8 candlestick patterns in 640x640 renderings of 20-candle daily charts.

The headline finding

Adding this detector's predictions to a plain price-feature baseline changed next-day direction accuracy by +0.0036 (95% bootstrap CI -0.0041 to +0.0114, McNemar p = 0.443) over 1936 out-of-sample days on SPY. The interval contains zero. Neither variant beats the majority-class rate of 0.5537, and both have a ROC AUC near 0.496.

So: this model detects the patterns well, and detecting them does not help predict the next day. That is the result, reported as found.

Labels are rule-based, not human-verified

Training boxes came from TA-Lib's CDLxxx functions. Nobody annotated these charts by hand, so this model imitates a published heuristic. Where TA-Lib and an experienced trader would disagree about the same candles, this model follows TA-Lib. Detection metrics below therefore measure agreement with a rule, not correctness.

Detection, held-out test split

class	instances	precision	recall	mAP@50
Hammer	557	0.665	0.720	0.757
ShootingStar	570	0.801	0.798	0.893
BullishEngulfing	1178	0.999	0.983	0.995
BearishEngulfing	1368	0.908	0.997	0.976
MorningStar	198	0.602	0.773	0.772
EveningStar	180	0.773	0.550	0.669
Doji	5261	0.893	0.948	0.975
Harami	2350	0.824	0.946	0.944
aggregate		0.808	0.839	0.872

Reported per class because the classes are very unevenly represented. The pattern in those numbers is not random: classes defined by a relationship between whole candle bodies (Engulfing, Harami) are close to solved, while those defined by a proportion inside a single candle (Hammer, Shooting Star, the Stars) are much harder -- at this resolution a candle body is only a few pixels tall, so the measurement the rule depends on is the one the image barely resolves.

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

model = YOLO(hf_hub_download("rohanjain2312/candlestick-pattern-recognition-system-yolo", "best.pt"))
results = model.predict("chart.png", conf=0.25)

Charts must be rendered the same way the training data was: 20 candles, 640x640, no axes, no gridlines, no volume panel. Use src/data_pipeline/render_charts.py from the repo -- a differently drawn chart is an input distribution this model has never seen. Note also that Ultralytics expects BGR arrays while most renderers emit RGB.

Splits

Chronological with an embargo gap, never random. Consecutive windows overlap in 19 of 20 candles, so the first 20 windows after each boundary are discarded and no test chart shares a candle with a training chart.

Not investment advice.

Downloads last month: 645