KILLED · pareidolia
Hypothesis
A convolutional net trained on rendered candlestick chart images learns profitable visual patterns end-to-end.
Math
$$ \hat y = \text{softmax}\big(W\cdot \text{CNN}(\text{image}(P_{t-k:t})) + b\big) $$
Method
Render OHLC windows to images, train a CNN classifier on forward-return labels, strict temporal split.
Results
| Validation accuracy | ~51% |
| Learned features | mostly recent-trend slope |
| vs a 2-feature logistic baseline | no improvement |
Rendering numbers as pixels and convolving them throws away precision and adds no information. The CNN rediscovers "recent slope" — reproducible with a two-line baseline — at far higher cost. Pareidolia with a GPU. Killed.
Turning structured numeric data into images to use a CNN is almost always a step backwards. The net can only learn what the pixels preserve, and pixels lose precision.