I frequently see ppl using Reinforcement Learning for trading single assets (e.g., BTC-USDT and I cannot wrap my head around why this could be superior than Supervised Learning. For instance, consider using the last 30 time steps to make a prediction if the price goes up or not. Supervised learning would give you this information way more efficiently than Reinforcement Learning. Reinforcement Learning needs to evaluate a value function that has a lot of overhead and is also less sample efficient and would also be different in terms of the objective function. However, both should basically work but I just don't see any advantage in Reinforcement Learning here, am I right or do I overlook something here? Please let me know if this is not understandable – I'd come up with way more explanations here if needed…
Moreover, many papers exist in port folio optimization….Like havin 5 Assets or so to let the Reinforcement Learning based agent optimize for the best split. This would make sense imho…Am I right with my considerations?
By the way: I'm not an algotrader; in general I also do not have too much experience in the world of trading/finance. But I've many years of professional experience in research-oriented Reinforcement Learning and also have some experience in other paradigms of ML and also classical AI methods. But I have been thinking about to just sit down for a couple of week and just give it a try… 🙂
Submitted November 08, 2020 at 09:57AM by -EniQma-