Another super-obvious way to spot a bad quantitive financial machine learning paper
Okay, I have to write a line here, so you don’t already see the trick when reading across the title of my article. :D
……………………………………………………………………………………….
Ready? Here it comes!
No transaction cost included. Guys, literally, hit ctrl + F and search for “cost”, “fee”, “rebate” or something like that. If you don’t find it, chances are high, you are about to be excited about a paper that shows you a great performing trading strategy, however will instead greatly disappoint you :D.
I know, this is so obvious, but, honestly, how many times do people forget about this, actually. There are TONS and TONS (or to say it in Trump’s words BILLIONS and BILLIONS) of bullcrap medium articles out there which try to sell you some kind of “indicator-optimized” strategy, which already is an inherently not working thing, but then on top, even without including any trading cost in their simulations. Unbelievable.
In my opinion, every scientific paper that includes a backtest of a trading strategy should have to at least include one version of that strategy with realistic fees. If fees are not contained, you can still try to infer the trading fees by dividing the total return of the strategy by the number of trades done over the backtest period, if these quantities are given. This will provide you with a maximum fee value that this trading strategy could withstand. But be sure to make it a very safe value of the fees (covering slippage and other market friction effects)
I will write down a more complete article about trading strategy simulation best practices, dealing for instance with in- and out-of-sample-simulation, etc. and post the link here later.