AAI_2025_Capstone_Chronicles_Combined

5.3) Time Series Stock Prediction

The final performance analysis was to try the capability of TurbaNet to train a swarm simultaneously on different data to learn different trends. Take for example the problem of predicting a stock price over time. This is a well studied problem and LSTMs tend to be an acceptable solution, so network architecture was used for this example. The challenge arises when predicting time series for dozens, hundreds, or thousands of individual stocks simultaneously. With PyTorch, one would have to train each model individually but with TurbaNet, the swarm can be trained concurrently leading to a massive performance boost. Figure 5.7 displays the distribution of the average errors across the testing data of the PyTorch and TurbaNet swarms for the 500 stocks that they were trained on. While the analysis here is largely qualitative, it provides useful insight into TurbaNet’s effectiveness and it shows that the TurbaNet models achieve relatively similar performance to the PyTorch models.

Figure 5.7 Average Model Error Across 500 Tickers

143