🛠️Model architecture
Last updated
Last updated
As mentioned earlier, the input data is fed into a convolutional layer where 1x54 filters are used to capture daily price patterns, where N represents the number of features. The convolutional neural network layer extracts trends and information by capturing and retaining the majority of the variance of the dependent variable. The convolutional operation within the model can be visualized in Figure 3.
The output data is not pooled to retain complex details as well as preserve the time aspect of the input, which is a 30-day lagged data for our analysis. This is primarily due to LSTM, a specific type of RNN, performing well on time series analysis models compared to other models. Furthermore, the data is passed through multiple LSTM layers. This allows the model to understand the underlying market trends not only based on the variables used to model a single instance but also on the data from previous days. Figure 4 depicts the loop architecture of LSTM as a fundamental component of the model.
The output from the LSTM layer is then passed through the Time Distributed layer (TDL). TDL is used to maintain the flexibility of the multivariate model. Currently, for the discussed experiments, the objective is to have input point predictions based on Solana (SOL/USDT). However, it is noted that for future experiments, we also want to predict Take Profit and Stop Loss (TP/SL). Therefore, in order to predict the price direction/entry price of Solana and TP/SL from the time distributed layer of this single model, it can also be used in the future without many changes in the model architecture. As this will also add an additional dimension to the input data, a 3D CNN will be used instead of the current 2D CNN that is being used as the initial layer for our model. This layer will perform a similar task of feature extraction based on the input features used on an additional dimension as well as for TP/SL. The complete model architecture is depicted in Figure 5.
