Stock movement and image classifier

Result: a stock movement baseline and an image classifier, with the leakage and baseline caveats written down. Scores are add both results.

Problem

These are two small projects. The stock task is easy to fake with a random split, and an image classifier is easy to report without a baseline. I keep them in the archive because they are useful breadth, not the strongest hiring signal.

Data

Stock data: add source, ticker universe, and date range. Image data: add dataset and class count. The stock split is walk forward. The image split is add split method.

Approach and why

Stock:

Build features only from information available before the prediction time.
Use a walk forward split instead of a random split.
Compare with a majority class baseline.
Report precision and recall, not only accuracy.

Image classifier:

Start with a simple baseline before a larger model.
Report a confusion matrix and per class metrics.
Check for duplicate or near duplicate images across the split.

The baseline and the split are the project. The model is the easy part.

Results

Task	Baseline	Model	Notes
Stock movement	Majority class add score	add score	Walk forward
Image classifier	add baseline	add score	add per class notes

Evaluation

Leakage: walk forward for stock, duplicate check for images, and a feature availability check.
Baseline: majority class for stock, simple classifier for images.
Validation: time ordered for stock, held out for images.
Metric: accuracy plus precision and recall for stock, per class metrics for images.
Reproducibility: fixed seeds, saved split dates, and a versioned dataset snapshot.

A reviewer should challenge any stock accuracy without a walk forward split and any image accuracy without a baseline.

Limitations

The stock result does not include trading costs, slippage, or a live paper trading period. The image classifier is a small scope exercise. Both projects are in the archive for a reason.

Next steps

Rebuild the stock task around one decision with a clear cost model. Add a calibration check. For the image classifier, add a stronger baseline and a data audit for duplicates.