Stock Signal Platform Using Screenshot
Budget: $10 – $30 USD
I need a complete, end-to-end pipeline that starts with a screenshot of a live trading terminal, pulls every ticker, quote, or indicator visible on the image, and ends with a single buy-or-sell probability score that consistently reaches 90 %+ confidence.
Scope
• Screenshot ingestion and parsing: an OCR/visual-recognition routine that can handle typical broker or data screens, isolate each symbol, price, volume and time stamp, and structure it as tabular data.
• Feature engineering: enrich the extracted snapshot with any historical data you deem necessary (you may tap your own CSV archives or free APIs), then create derived variables such as intraday return, volatility, or technical signals.
• Probabilistic modeling: build and train a machine-learning model—choice of algorithm is open, as long as it is explainable and achieves the target accuracy on held-out data. Neural networks, tree ensembles or hybrid methods are all acceptable if they outperform simpler baselines.
• Scoring interface: a lightweight CLI or web endpoint that accepts a new screenshot and returns a JSON payload containing “BUY”, “SELL” and the associated probability.
• Documentation & reproducibility: clean, commented Python code (TensorFlow, PyTorch, scikit-learn or similar), requirements.txt, and a short read-me describing set-up and retraining steps.
Acceptance criteria
1. End-to-end demo with two unseen screenshots producing valid scores.
2. Back-test report showing at least 90 % directional accuracy across 1,000+ historical samples.
3. Code passes linting and runs on a fresh environment with the provided instructions.
If parts of the stack—OCR, data augmentation, or model selection—require iterations, please outline your milestone plan so we can agree on checkpoints and validation tests before moving forward.
Scope
• Screenshot ingestion and parsing: an OCR/visual-recognition routine that can handle typical broker or data screens, isolate each symbol, price, volume and time stamp, and structure it as tabular data.
• Feature engineering: enrich the extracted snapshot with any historical data you deem necessary (you may tap your own CSV archives or free APIs), then create derived variables such as intraday return, volatility, or technical signals.
• Probabilistic modeling: build and train a machine-learning model—choice of algorithm is open, as long as it is explainable and achieves the target accuracy on held-out data. Neural networks, tree ensembles or hybrid methods are all acceptable if they outperform simpler baselines.
• Scoring interface: a lightweight CLI or web endpoint that accepts a new screenshot and returns a JSON payload containing “BUY”, “SELL” and the associated probability.
• Documentation & reproducibility: clean, commented Python code (TensorFlow, PyTorch, scikit-learn or similar), requirements.txt, and a short read-me describing set-up and retraining steps.
Acceptance criteria
1. End-to-end demo with two unseen screenshots producing valid scores.
2. Back-test report showing at least 90 % directional accuracy across 1,000+ historical samples.
3. Code passes linting and runs on a fresh environment with the provided instructions.
If parts of the stack—OCR, data augmentation, or model selection—require iterations, please outline your milestone plan so we can agree on checkpoints and validation tests before moving forward.