What kind of returns should I use for my model?

Question

I'm building a machine learning model with the aim of learning a daily strategy of buy or sell the stock.

I was wondering if I should use adjusted close price or something else to calculate returns (I was thinking about considering open price/close price the day before) and to evaluate the strategy. I know that adjusted prices offer a better representation of the price as they account for dividends and other things, but with in that way the results I would get are they consistent with reality? In other words, I fear it may happen that a strategy which is successful for returns calculated through adjusted prices would be not the same in a realistic world.

Answer 1

You can use the adjusted close price; it is far better than using unadjusted prices and having your strategy tell you to short a stock on its ex-dividend days.

The bigger issue is using closing prices -- adjusted or unadjusted. Closing prices are determined by an auction and the presence of your order in the auction will change the auction price. Furthermore, you do not know what the auction price will be until the auction is over. Many supposed trading strategies falter on this issue.

Answer 2

Tautolotigally, a stock holder receives the dividend if they're the stock holder of record on the record date. Even if your trading strategy assumes that you will put on some position at the beginning of the day and then always flatten before the close, keeping no risk overnight, whatever information is contained in the return series assumes that somebody will pay for receiving the economic benefit of the dividend at the ex date close. The price-only return between the ex date close and the next day open shows a loss equal to the dividend and is just meaningless. You should include the dividends in your returns, even if you never receive them using your trading strategy.