What is a convolutional neural network and does it have finance applications?

Question

AcadiFi · Accepted Answer

A convolutional neural network (CNN) applies learnable filters (kernels) that slide across input data and detect local patterns. Three key layers:

1. **Convolution**: each filter produces a feature map by convolving with inputs. Captures local correlations (e.g., edges in images, patterns in time series).
2. **Pooling**: downsamples feature maps (max pool, average pool) to reduce dimensionality and provide translation invariance.
3. **Fully connected**: classic dense layer at the end for classification or regression.

Finance applications:
- **Price chart analysis**: treat a candlestick chart as an image; CNNs detect patterns like head-and-shoulders, double bottoms. Research (Sezer et al., 2018; Jiang, 2020) shows CNNs applied to Gramian angular fields of returns can outperform basic technical indicators.
- **Satellite imagery**: count cars in Walmart parking lots, estimate oil storage tanks from infrared imagery, forecast crop yields from NDVI maps. Orbital Insight and RS Metrics use CNNs to sell alt-data signals.
- **Document processing**: OCR of earnings releases, extraction of key financial figures from 10-K filings.
- **Time series as 1D convolution**: input a sequence of OHLCV bars; 1D convolutions detect momentum and reversal patterns.

Example architecture for 60-day return prediction from 1D OHLCV series: Conv1D(32 filters, kernel 5) -> ReLU -> MaxPool(2) -> Conv1D(64, 3) -> ReLU -> Flatten -> Dense(16) -> Dense(1 regression).

```mermaid
graph LR
  A[Input 60x5] --> B[Conv1D]
  B --> C[Pooling]
  C --> D[Conv1D]
  D --> E[Flatten]
  E --> F[Dense]
  F --> G[Return forecast]
```

For CFA: know the intuition that CNNs exploit spatial or temporal locality, trainable filters, and weight sharing - fewer parameters than a fully connected net of comparable capacity.

What is a convolutional neural network and does it have finance applications?

Master Level II with our CFA Course

Related Questions

Practice Questions