I have a Neural Network model that provides predictions for the future returns of a portfolio comprising stocks and cryptocurrencies. The original model operates on standard time bars and generates daily predictions at a specific time, utilising processed OHLC (Open, High, Low, Close) candlestick data as input.
My question is general and the details of the model are not important. The most important thing is that the features are computed from the time bars. For example, the input dataset can be the intraday 5min candles for each instrument, where some processing is involved on the values of the candles.
I would like to extend this Neural Network model to make it work with dollar bars, instead of time bars. The first questions that comes to my mind are
- The original model is time-dependent, issuing predictions daily at specific times. When transitioning to dollar bars, this time sensitivity is lost. How can we establish criteria for determining when the model should make predictions? For a single-instrument model, one possible approach might involve assessing the average daily count of dollar bars within a given sampling size and making predictions accordingly. However, when dealing with a multi-instrument model, how can we adapt this criterion?
- What are the most prominent features that can be derived from dollar bars and are suitable for serving as inputs to a neural network? In the earlier example, the input dataset consists of a dataframe containing 5-minute intraday candles. We can reliably anticipate that there will be a 10:00 AM candle for any instrument every day. Consequently, by choosing this candle as a feature, we ensure that the input dataset consistently contains information about the 10:00 AM candle. However, when dealing with dollar bars, there is no assurance that two instruments will share the same dollar bar. So, it seems there is an ambiguity on how I can univocally define the input features for a Neural Network model.