News for hackers. Stuff that matters.
Re:Ask HN: How credible is Reinforcement learning in finance... by shoo on Wednesday September 02, 2020 @01:29PM (#24355564)
> However there is no way to simulate the price

Compared to other real world problem domains which are far less computerised it seems relatively simple to get fresh data for a trading system: let the system execute actual trades and measure the actual response.

It might be difficult to do this without having a budget to burn in running experiments, both in terms of money that is going to be lost while the system makes poor trades, setting up controls to ensure that not too much money can be lost in any experiment, and subscribing to data feeds that give a richer view of how the market responds (e.g. order book info).

In contrast, consider trying to get fresh data in materials science etc where you may need to manufacture small batches of materials and then test them. It might cost $10k in materials and weeks of work with expensive machinery and skilled technicians to generate a dozen or so fresh data points.

There are no comments on this comment yet.