AI Weather Forecast & RNN model
Reflection on DeepMind's GraphicCast
I've always been into weather, much to my wife's teasing about the usefulness of knowing the weather days in advance. But I see weather patterns as a way to spot arbitrage opportunities. As a trader, quickly identifying these opportunities isn't just about making more money; it's also about improving market efficiency with better information. I believe advanced machine learning models can help uncover patterns in the factors influencing weather.
Recurrent Neural Networks (RNN) caught my eye for handling correlated datasets in machine learning. Unlike traditional cross-sectional data, time series data emphasizes the order of observations, showing a high correlation between them. This makes assuming an Independently Identically Distributed (IID) scenario challenging in time series modeling. RNN, needing sequential data, adds an extra dimension compared to simple Deep Neural Network (DNN) models, highlighting the importance of maintaining temporal order.
When I first encountered RNNs, I was puzzled by how the model could predict when the input sequence length varied. Unlike DNN models, where input data is typically fixed, RNNs produce predictions of whether the input sequence is one or ten, as long as the features of each observation remain constant. This aspect intrigued me, especially in a common RNN example predicting the next character based on Shakespearean plots, during my Deep Learning course.
Understanding the role of hidden layers, a commonality with other deep learning models, was the breakthrough. The RNN's input layer unfolds features like a form of feature engineering, enabling hidden layers to memorize crucial aspects of the input and understand relationships between features. This mechanism explains why, even with just one character like 'a,' the model predicts subsequent characters based on a probabilistic model.
Building on my interest in weather forecasting and the potential of RNN models for identifying arbitrage opportunities in the financial market, I applied my knowledge to create RNN models for predicting future energy prices. In an MBA course project, I focused on developing a local electricity price forecasting model. Despite limited access to data, the RNN-based model performed well, outperforming other statistically-based models in time-series predictions.
My enthusiasm for machine learning and its applications in weather forecasting deepened when I came across an article from the Google DeepMind team discussing GraphicCast. It seems to be an RNN variant, delivering predictions at six-hour intervals. This prompted me to document my understanding of RNNs and how machine learning can surpass traditional forecasting methods.
GraphicCast's notable distinction is its "model-free" nature compared to traditional models. Traditional models rely on meteorological theories based on the mathematical deployment of physical theories. The flexibility of NN models, unbound by strict physical principles, stands out as a strength in contrast to traditional approaches.
While ML-based weather forecast models have pros and cons, a controversial topic revolves around their predictability in unexpected events. My initial concern was that ML models might struggle to explain the 'why' behind predictions. However, I view this as an opportunity for further exploration. Machine learning's primary goal is not necessarily to understand but to enhance predictive capabilities, and I aim to focus on its prowess in forecasting weather events.
In practical terms, I anticipate ML-based weather forecasts making a substantial impact on businesses, particularly in agriculture and commodity markets. The susceptibility of these markets to inclement weather forecasts makes them ideal candidates for leveraging advanced ML models. Despite these promising developments, it's crucial to acknowledge the challenge of insufficient information, especially when working with proprietary or limited datasets.
As a takeaway from the DeepMind article and this reflection, my algorithmic trading bot currently relies on an RNN model trained through reinforcement learning, utilizing market price data and technical analysis. Looking ahead, incorporating signals from a 'weather watchdog' bot promises to enhance the efficiency of my trading bot, offering a comprehensive approach that factors in both market dynamics and weather forecasts. However, dealing with insufficient information remains an ongoing challenge, requiring continuous efforts to improve the robustness of predictive models.