Variance vs EV: Thinking in the Long Run
A +EV all-in that loses was still correct, and a bad call that wins was still bad. Learn how often favorites actually lose, what sample sizes can and can't tell you, and how to grade decisions instead of results.
Assumptions: All examples assume 100bb effective stacks in a 6-max online cash game with no rake, unless a different stack size is stated in the example.
You get kings in preflop against ace-seven, the ace bangs the river, and a buy-in slides across the table. Two different things just happened, and the entire skill of this lesson is keeping them apart: the decision earned you about 42 big blinds the moment the money went in, and the result took 100 from you a few seconds later. The decision was yours; the result belonged to the deck.
Players who can't separate those two unlearn good poker one bad beat at a time. They stop value-betting thin because "he always has it," start limping kings because "I always lose big pots with them," and quit bluffing because the last two got called. Results-oriented thinking is the main way beginners coach themselves into becoming losing players. The antidote is knowing — with numbers — how loud the noise actually is.
The "80/20" that loses three times in ten
Table talk calls this kind of spot "an 80/20." Run the actual simulation: K♠K♥ against A♦7♣ is 71% — kings lose this confrontation 29% of the time, not 20%. (A genuine 80/20 is something like kings against sevens, which computes to 80%.) The gap between the number people feel and the number that's true is itself a lesson: intuition systematically understates how often favorites lose.
The EV of getting 100bb in here, with 200bb in the final pot:
EV = 0.711 × 200 − 100 = +42bb
Now the part your brain will fight you on. A 29% event is not rare. Roll the spot ten times and the binomial math says you lose three or more of them about 59% of the time — losing a cluster of these all-ins isn't a cooler streak, it's the expected texture of the results. You will, with full mathematical regularity, watch this +42bb decision torch buy-ins in bunches and feel like the game is broken. It isn't. Nothing in those losses contains information about the decision. The +42bb was banked at the click; the deck just decides who holds the money short-term.
How often does an edge even show up?
Make it more general. Suppose you have a true 60/40 edge in some repeated confrontation — a big edge by poker standards. How often do you have a losing record anyway?
- Over 5 trials: you win 2 or fewer about 32% of the time
- Over 20 trials: you win 9 or fewer about 13% of the time
- Over 100 trials: you win 49 or fewer about 1.7% of the time
A third of the time, a clear favorite walks away from a short session as the loser. Even across 20 repetitions — more identical all-ins than you might see in a week — the edge fails to show up in the win column once in eight. Only out past a hundred trials does the truth become hard to miss. The long run is real, and it is long: edges are invisible at the scale of a session, faint at the scale of a week, and undeniable only at scales your emotions never operate on.
What your win rate can't tell you (yet)
The same logic applies to the biggest number you'll ever obsess over: your win rate. Take a genuinely good player — a true 5bb/100 winner with a standard deviation of 80bb/100, typical for 6-max cash. Run the variance math at three sample sizes:
- 500 hands (one evening): expected profit 25bb, but the 95% confidence interval on the observed win rate spans −65 to +75 bb/100, and the probability the session just loses outright is 44%. A single session contains essentially zero information about your skill. Hundreds of hands tell you almost nothing.
- 10,000 hands (a serious month for a casual player): the interval is still −11 to +21 bb/100 — you can't yet distinguish a solid winner from a slight loser — and a winning player still posts a losing 10k stretch 27% of the time.
- 100,000 hands: the interval finally tightens to roughly 0 to +10 bb/100, and the chance a true winner shows a loss falls to about 2%. Now the data is starting to talk.
Internalize the asymmetry: a downswing over thousands of hands is fully consistent with being a winner, and a heater over thousands of hands is fully consistent with being a loser. The player typing "I'm crushing, 30bb/100 over the weekend" and the player typing "I must be terrible, down 8 buy-ins this month" are making the same statistical mistake in opposite directions.
Grade the decision, not the deck
If results can't grade your play on any timescale you actually live in, what does? The numbers that were available at the moment of the decision. That's a discipline, and it works like this:
The big blind chases a gutshot for his stack: four outs is 9% on the river card, the price demands 43%, and the call is worth about −40bb the moment the chips go in. The 9♦ arrives; he drags 116bb. Was the call good? It cost him 40bb. The pot he's stacking is variance; the −40bb is forever. If you were the player holding the aces, the same separation applies in reverse: you lost a huge pot and made money on the hand in the only sense that compounds.
So build the habit explicitly:
- Re-grade hands by information at decision time. "Given his range and the price, was my number right?" If yes, the hand goes in the win column regardless of the river. If no, it's a leak to fix even if you dragged the pot.
- Review the losses you were supposed to take. When the 29% side wins, there is no lesson — and force-extracting one ("I should have folded kings preflop!") is how variance teaches you fiction.
- Catch result-creep in real time. "He's hit every draw against me, so I'm not value betting this" is variance rewriting your strategy. The fish who called three bad gutshots and hit two of them is more profitable to value bet, not less.
- Treat ten hands as zero evidence. A 10-hand sample of anything — your all-ins, your bluffs, your session — answers no question. You already know what ten trials of a 71% favorite look like: a coin-flip-ish chance of an ugly record.
What's signal, what's noise
None of this means you should never worry about losing. It means you should know where to look when you do. Results are noise on every timescale your emotions can feel — but decisions are signal on a sample size of one. The productive response to a losing week is not "what is my win rate now?" (the confidence intervals above just told you that number is mush). It's an audit of choices:
- Pull up the big pots and re-run the math from this module. Was the price right? Was the fold-equity estimate defensible? Would you make the same play again, knowing only what you knew then? A losing week where every answer is "yes" requires exactly zero changes.
- Watch for tilt leakage, which is diagnosable from tiny samples because it shows up in decisions, not results: the 3am session extension while stuck, the revenge 3-bet against the player who sucked out, the "I never get there anyway" fold with the right price. One clearly tilted decision is real signal. Ten lost flips are none.
- Be suspicious of lessons that arrive attached to pain. "Stop overvaluing one pair in 3-bet pots" discovered through calm review is probably real. "Never play AK, it always misses" discovered while staring at a four-buy-in hole is variance dictating your strategy.
One more piece of long-run hygiene: session boundaries are accounting fiction. Quitting "to lock up a win" or grinding past exhaustion "to get unstuck" both treat the session as a unit the deck cares about. It doesn't. Your poker life is one long session interrupted by sleep, and every hand is played at the EV available in that moment. The only legitimate in-session reasons to quit are the ones that change your future EV: fatigue, tilt, a lineup that got worse. "I'm up two buy-ins" changes nothing about the next hand's expectation — though if being up makes you play scared, that's a fatigue-class reason, and it's about your decisions, not the score.
The practical answer to variance: bankroll
You can't reduce variance — it's priced into the same aggression that generates your EV — but you can make it survivable. That's all a bankroll is: enough buy-ins behind you that the inevitable 27%-style losing stretches can't force you to quit or move down at the bottom of a swing. For the 5bb/100 profile above, the variance math suggests roughly 1,900bb for a 5% risk of ruin and about 2,900bb for 1% — call it 20 to 30 buy-ins as a working floor, more if your edge is smaller or your game is wilder. The buy-in formulas and moving-up rules belong to a later module; the concept belongs here: bankroll is what buys your +EV decisions enough trials to become money.
The long run isn't a place you arrive at. It's a way of scoring the game while you play it: every decision banked at its EV, every result filed under noise, and the discipline to keep your strategy pointed at the first number while the second one screams in your face.