The reason is pretty simple - probability of events are an important input to calculation of expected values. If the probabilities are off, expected value calculations will differ. Also, Expected Value works under the “law of large numbers” assumptions. That in turn brings into picture the “sequence of return” risk. Two drastically different sequences can lead to the same Expected Value but can have serious short term implications.

“The market can remain irrational longer than you can remain solvent.”

> Also, Expected Value works under the “law of large numbers” assumptions.

Technically, EV has nothing to do with sample size. But I get your point that in sufficiently small sample sizes and/or sufficiently large bet sizes you might need to think about utility rather than expectation.

"easy for casino table games"

I don't believe EV is easier to estimate in poker than trading. You need to estimate hand range and the consequences of actions later in the hand. It's extremely complicated.

Yes...bet sizing is something I worry about a lot, but it isn't clear to me how to apply the kelly criterion to a game where the risk/reward is mostly unknown and only based on a hunch.

You could make a table where row is guessed probability(input to Kelly criterion) and column it's actual probability and cells contain expected value and expected log value. And pick something which looks reasonable.

One better-than-nothing way is to look at historical data:

Looking at the closing price for each trading day, count how many times the stock ended higher and how many times it ended lower than the previous day.

Then you have your odds.