nnstock is a small prototyped project for storing stock market trends using neural network compression. It showcases how a neural network can be used to store data efficiently, but with less resolution.
10 years (2010-2020) of daily price data for different stocks from the S&P500:
3 Stocks:
Torch model size on disk: 77.50 KB
Raw data size in memory: 176.91 KB
10 Stocks:
Torch model size on disk: 77.50 KB (same as 3)
Raw data size in memory: 589.69 KB
This means, using the simple linear model here, the whole S&P500 10 year price data could be stored about 300 times more efficient.
| Ticker | Max % Error | Mean % Error | Median % Error |
|---|---|---|---|
| AAPL | 51.66 | 10.66 | 6.97 |
| MSFT | 25.15 | 6.03 | 4.03 |
| GOOG | 41.54 | 8.38 | 5.72 |
| MCD | 14.05 | 2.87 | 2.49 |
| NFLX | 76.50 | 9.88 | 6.08 |
| WMT | 22.00 | 5.85 | 4.88 |
| JPM | 28.13 | 4.00 | 2.87 |
| COST | 18.85 | 2.95 | 2.49 |
| IBM | 20.58 | 3.18 | 2.50 |
| DIS | 20.71 | 3.35 | 2.76 |
Model prediction time: 475.17 µs (CPU)
Raw data lookup time: 45.78 µs
The model lookup is much slower in this case, but could be signicantly improved with proper inference server.
PyTorch, yfinance, matplotlib, numpy, sklearn