citadel's approach to researching a signal starts with a question almost nobody asks before backtesting. max bane, quantitative research lead there: "say we wanted to research some kind of momentum signal. why should something like momentum exist at all? why should it even arise? aren't markets supposed to be efficient?" "if we're going to hypothesise momentum, we should have in mind some kind of causal mechanism for why it should happen." two he offers: information propagates at finite speed. not everybody sees the news at the same time, so they act in sequence, and that creates autocorrelation in the price changes. or: someone with a large position splits it into smaller trades to control their transaction costs, and each one pushes price a little. here's why it matter: "the reason to go through this exercise of thinking through what the causal mechanisms are is that it gives you something to latch on to in terms of how you would test your hypothesis." if it's news propagation, go get data on how many people follow each stock. the theory predicts momentum is stronger where fewer people are watching. if it's institutional order splitting, quantify institutional versus retail interest. the theory predicts stronger trends in one than the other. "what is my theory, what predictions does it make, and how can i try to disprove the theory?"