Trade State Classifier: Open vs Not-Open Position Detection

A fine-tuned YOLO26n-cls model that classifies livestream trading screen captures as showing an open position or no open position. Designed for real-time mirroring of trades shown on trading livestreams (TradingView, TopstepX, Tradovate, and similar platforms).

What it does

Given a screenshot (or video frame) from a futures trading livestream, the model answers one binary question: is there a filled/current position visible on screen right now?

Input:  1280×720 screenshot from a trading livestream
Output: open (position visible) | not_open (no position)

This is the state-gate component of a larger pipeline that:

Detects position state transitions (flat → open → flat)
Extracts trade details (direction, entry, SL, TP, quantity) at the transition moment
Emits validated signals to Discord → NinjaTrader 8 for execution

Training approach

Base model: YOLO26n-cls (Ultralytics, AGPL-3.0)
Fine-tuning data: ~1,700 owner-verified and confluence-confirmed crops from screen captures the author collected directly, spanning two independent channels and two UI layouts (a TradingView-style and a TopstepX-style interface)
Training: 20 epochs, 512×512, batch 16
Labeling pipeline: Automated Whisper speech-cue extraction → local VLM (qwen2.5vl:7b) keyframe labeling → speech+visual confluence confirmation → dense segment extraction around confirmed trade windows

Performance

Metric	Result
Temporal validation accuracy	100% (341/341)
Held-out false positives	0/165
Entry detection latency	+2 seconds after fill
Exit detection latency	+10 seconds after spoken exit
Inference speed	~15ms/frame on CPU

Usage

from ultralytics import YOLO
import cv2

model = YOLO("best.pt")
image = cv2.imread("screenshot.png")
results = model.predict(image, imgsz=512, device="cpu")
class_name = model.names[int(results[0].probs.top1)]
confidence = float(results[0].probs.top1conf)
print(f"{class_name} ({confidence:.3f})")
# "open (0.950)" or "not_open (0.980)"

With ROI cropping

For streamers whose chart occupies a known region, crop first:

roi = (0.0, 0.08, 0.78, 0.82)  # normalized x,y,w,h
h, w = image.shape[:2]
crop = image[int(roi[1]*h):int((roi[1]+roi[3])*h),
             int(roi[0]*w):int((roi[0]+roi[2])*w)]
results = model.predict(crop, imgsz=512, device="cpu")

Important note on test data

This model was trained on real trading-platform screen captures and learns platform-specific visual features (position badges, PnL overlays, order tool rendering). It does not generalize to simplified synthetic charts or mock screenshots — it requires real platform UI captures to function correctly. No test images are included to protect the privacy of everyone involved; see the GitHub repo for the full evaluation pipeline.

Limitations

Visual-only: The model cannot detect a position if the position badge/box is not visible on screen (e.g., the streamer switched to a different chart mid-trade).
Platform-specific appearance: A position box on TradingView looks different from TopstepX. This model was trained on both but may need additional data for other platforms.
Hypothetical vs filled: A drawn position-tool box (hypothetical trade plan) looks identical to a filled position. This model detects the visual state; confirming an actual fill requires the full pipeline (speech confluence + state transition logic).
GEX charts: Gamma exposure screens with green/red bars can produce false positives. Use platform-anchor templates to gate classification to the execution platform screen only.

Ethics & privacy

Training data consists of screen captures collected directly by the author, handled with respect for the rights and privacy of everyone involved; no raw frames are distributed
No raw training images are included in this repository
The model should not be used to copy trades without the knowledge or consent of the person presenting them
This model is a research prototype, not financial advice

License

AGPL-3.0 (inherited from the Ultralytics YOLO base model).

Reaver1000
/

trade-state-classifier