ChartVision

Candlestick chart pattern recognition with a CNN, plus Grad-CAM heatmaps that show where on the chart the network is looking.

Real OHLC data (yfinance) is labelled with geometric rules, rendered into minimal candlestick images, and classified into one of five pattern classes. Two models are compared: a small from-scratch CNN and a ResNet18 with transfer learning.

Classes

Class	Family	Bias
head_and_shoulders	reversal	bearish
double_top	reversal	bearish
double_bottom	reversal	bullish
bull_flag	continuation	bullish
no_pattern	negative	n/a

Repository layout

code/      source: data pipeline, models, training, evaluation, Grad-CAM, tests
figures/   figures used in the report and slides
report/    final report (PDF)
slides/    final presentation (PDF)

code/

src/
  data/       fetch_ohlc, rule_labelers, render_charts, build_dataset, dataset
  models/     baseline_cnn, resnet18
  training/   train, evaluate, compare_models, gradcam
  inference/  prediction helpers
  utils/      config, seeding
configs/      default.yaml: single source of truth (paths, tickers, seeds, hyper-params)
tests/        pytest unit tests for the rule labelers
make_*.py     figure / slide-asset generation scripts

Raw/processed data, model checkpoints, and generated figures are not committed; they are reproduced by the pipeline.

Setup

cd code
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install torch torchvision

Reproduce

python -m src.data.build_dataset
python -m src.training.train     --model resnet18
python -m src.training.evaluate  --checkpoint checkpoints/resnet18_best.pt

Tests

pytest

---

Computer Vision Short Project, Master in Innovation and Research in Informatics, FIB UPC.