Hi-lo guide
Koi Hi-lo: when card memory helps, and when it cannot
"Koi" names this page and has no link to the content. The content is a practical question about hi-lo: if I remember which cards have already appeared, can I call better?
The short answer
It depends entirely on how the next card is made. If each card is generated fresh, for example from a hash turned into one of 13 ranks, memory is worthless: earlier cards have no influence at all. If the cards come from one real deck that is not reshuffled, earlier cards do change what is left, and memory can shift the chances. Even then it only helps if the payouts do not adjust to the cards that are gone.
A worked deck where memory matters
Suppose the game deals from one 52-card deck, the face card is an 8, and you remember that seven high cards have already been dealt: a 9, a 10, a J, a Q, two kings and another Q. With fewer high cards left, lower becomes more likely than the 13-rank model says.
| Deck model | P(higher) | P(lower) |
|---|---|---|
| Infinite deck (13 ranks) | 38.46% | 53.85% |
| Fresh 52-card deck | 39.22% | 54.90% |
| 52 cards minus 7 seen | 29.55% | 63.64% |
In the endless model, lower from an 8 wins 53.85%. In this depleted deck it wins 63.64%. If the game still pays the 13-rank price of 1.84x, the expected return of that call is 1.1700 per unit, which is above 1. In that narrow case, memory is worth something real.
The same cards with the other call
The flip side is just as large. Higher from the same 8 drops from 38.46% to 29.55%, so a player who ignores the seen cards and calls higher is worse off than the stated edge suggests.
When memory cannot help
- Hash-generated cards. Each round maps its own random number to a rank. No card is removed, so nothing is depleted.
- Reshuffle every round. If the deck is rebuilt before each deal, the seen cards are back in.
- Live repricing. If the game recomputes each multiplier from the cards that remain, any shift in chance is matched by a shift in price, and the return stays at 1 - e.
- Early in the deck. With only a few cards gone, the shift is too small to beat a typical edge.
The last point is easy to miss. Three or four seen cards move the chances by about a percentage point, which is the same size as a 1% edge. Real memory gains need a long, uneven run of removed cards, and they only exist when the deck model and the price model disagree.
Is this the same as counting cards at blackjack?
The idea is related: depleted decks change the odds. The details differ, and in hi-lo the effect only exists with a real finite deck priced as if it were infinite.
Can memory help with generated cards if I watch long enough?
No. A rank produced from a fresh random number has a 1/13 chance for every rank, whatever came before.