There is a slew of academic papers that look at basically the same base model while switching optimization algorithms for either its hyperparameter or model parameter tuning. While I acknowledge the urgency of a forecast for day ahead, week ahead and month ahead forecast, is convergence a game changer for 5-10 year ahead forecasts?
The prediction of stock groups values has always been attractive and challenging for shareholders due to its inherent dynamics, non-linearity, and complex nature.
I would rather say that inference time is the more important factor, depending on the neccessity to re-train your model. If you have constant concept drifts this might be necessary, but on the other hand you want a stable, reliable model that generalizes well. having said that, long-range predictions for the finacial market (e.g. stock price prediction) is not really feasible in terms of 5-10 years due to the inherent dynamic. That is why short-term-models that predict values that can lead to decisions have to be produce fast and the decision has to be executed immediately (e.g. in the stock market the majority is traded algorithmicly). So, the the internet connection and how fast your model can infer and execute the trade, is the biggest game changer probably.
Thank You, Khan and Heinrich for answering the question. It is interesting to see perspectives from folks involved in financial forecasting. Hope people from energy domain particularly electricity load forecasters could also share their perspectives.
- Badges
- Science topic