PROBLEMS OF USING NEURAL NETWORKS TO PREDICT THE PRICE OF VIRTUAL ASSETS

Background. Predicting the prices of virtual assets is an important task due to their high volatility. Neural networks are widely used for such tasks, but often face the problem of naive predictions, when the next value is too similar to the previous one, which reduces the forecasting efficiency. Materials and Methods. The study was conducted on the basis of Bitcoin price dynamics data (Coinbase exchange) for the period of May 4, 2021 - April 9, 2024. Two methods of data normalization were considered: linear (Min-Max) and ratio normalization, which ensures data stationarity. A neural network with bidirectional LSTM layers was trained on 10 previous values to predict one subsequent value. The TensorFlow library, in particular the Keras API, was used for training process with customizable parameters: 150 epochs, 32 batch size, Adam optimizer, and MSE loss function. Results and Discussion. Linear normalization showed the worst results, as the model loses its ability to predict if future values fall outside the training set. On the contrary, ratio normalization showed much better results, allowing the model to take into account the dynamics of changes even beyond the minimum and maximum values of the training data. Using an additional multiplier for the normalized data (k = 400) improved the results for the training data, although the improvement was not significant on the test set. Adding additional price characteristics (at the opening, closing, and minimum price during this period) failed to eliminate the problem of forecast bias. The analysis of the results showed that the model is often based on the repetition of previous values, which indicates its limitations in capturing complex relationships in the data. Conclusion. The study confirms that neural networks have limitations in the task of predicting virtual asset prices. The problem of naïve prediction is a key obstacle that limits the effectiveness of models. The use of ratio normalization with a k-factor improves accuracy, but it is not enough to solve the main problem. New approaches or improvements to existing methods are needed to allow the model to better understand complex relationships in time series and reliably predict future changes.