Understanding Poker Variance: Standard Deviation, Confidence Intervals, and Sample Size

Your tracker reads +4.2 bb/100 over 18,000 hands at $1/$2. A reg in your chat says you're printing. Your gut says the last week felt terrible. Which number do you trust?

None of them—not yet. 18k hands is a short sample for poker statistics. The number on screen is a measurement, not proof of skill. Variance is the gap between that measurement and your true long-run win rate. Until you know how much noise surrounds the signal, you cannot tell skill from a hot week.

Primedope, PokerCharts, and hhDealer all describe the same framework: poker results are random variables; standard deviation (SD) quantifies the noise floor; confidence intervals (CI) bound where your true win rate probably lives; and sample size controls how tight those bounds get—by the square root of hands played, not linearly.

Why Outcome Graphs Are Noisy

Amateur players treat a bankroll curve like a verdict. Professionals treat it like a single sample—one noisy snapshot of an underlying win rate you cannot see directly.

Term	Poker meaning
True win rate (μ)	Your long-run bb/100 if decisions stayed frozen forever
Observed win rate	Tracker bb/100 over your sample
Standard deviation (σ)	Typical swing width per 100 hands (bb/100)
Sample size (N)	Hands logged
Confidence interval	Range that likely contains μ

Variance is not "bad luck" as a moral category. It is statistical dispersion: even perfect +EV play produces wildly different short-run totals because hand outcomes are correlated random events with heavy tails (coolers, multiway pots, all-ins).

Standard Deviation as the Noise Floor

Standard deviation answers one question: how wide do my results swing around my average, per 100 hands? Tracker software reports it as Std Dev bb/100 (PokerTracker, Hold'em Manager) or you can estimate from industry tables when data is thin.

Typical online NLHE SD ranges (Primedope, PokerCharts, The Poker Bank):

Format / style	Typical σ (bb/100)	Why
Full-ring 9-max	60–80	Fewer multiway all-ins
6-max cash	85–110	More aggression, bigger pots
LAG / multi-table 6-max	100–120	Volume + wider ranges
PLO 6-max	120–160	Pot geometry amplifies swings
Live full-ring	80–120	Loose, multiway, slower pace

Higher σ = noisier results. A 5 bb/100 winner with σ = 110 looks identical to a breakeven player over 20k hands far more often than either would guess. Skill sets the center of the distribution; SD sets how long you must play before the center becomes visible.

Rule of thumb over N hands: cumulative result dispersion scales as:

$\text{Total SD in bb} = \sigma \times \sqrt{N/100}$

Example: σ = 90, N = 50,000 → total 1-SD band = 90 × √500 ≈ ±2,012 bb. At $1/$2 (100 bb = $200), that is roughly ±20 buy-ins of swing around expectation—before asking whether your measured win rate is even accurate.

Confidence Intervals (The Error Bar)

Your observed win rate is an estimate of μ. A 95% confidence interval is the error bar: if you could replay your career thousands of times and collect samples of N hands each, 95% of those measured win rates would fall inside this band.

Standard formula (large samples, normal approximation):

$\text{95% CI} = \text{WR} \pm 1.96 \times \frac{\sigma}{\sqrt{N/100}}$

Where WR and σ are in bb/100, N is hands, and 1.96 is the z-score for 95% confidence (PokerCharts, gamblingcalc.com, hhDealer).

Worked example: WR = +5 bb/100, σ = 90, three sample lengths:

Hands (N)	Standard error	95% CI	Plain English
10,000	±9.0 bb/100	−12.6 to +22.6	Almost anything still plausible
50,000	±4.0 bb/100	−2.9 to +12.9	Direction hint only
100,000	±2.8 bb/100	−0.6 to +10.6	Maybe winning; still wide
250,000	±1.8 bb/100	+1.5 to +8.5	First band that excludes zero
500,000	±1.2 bb/100	+2.5 to +7.5	High confidence you are +

The brutal insight: 50k hands feels like a career chunk online, yet a +5 bb/100 screen could still be a −3 bb/100 breakeven reg. The square-root law is unforgiving—you need 4× the hands to halve the error bar, not 2×.

Quick intuition without full CI: ±1 standard error (σ/√(N/100)) covers roughly 68% of outcomes; ±2 SE covers roughly 95%.

Sample Size: How Many Hands Do You Need?

Reverse the CI formula when planning a study goal: how many hands to pin μ within ±X bb/100 at 95% confidence?

$N = \left(\frac{1.96 \times \sigma}{X}\right)^2 \times 100$

For σ = 90 bb/100:

Target precision (±)	Hands required	Calendar reality (online 6-max, ~500 hands/hr)
±5 bb/100	~125,000	~250 hours
±3 bb/100	~346,000	~690 hours
±2 bb/100	~778,000	~1,550 hours

Lower true win rates need even more volume because the CI must clear zero, not just hit a precision target. A 2 bb/100 grinder at σ = 90 needs roughly 780,000 hands before the 95% CI lower bound turns positive—versus about 125,000 for a 5 bb/100 winner at the same σ. Same noise floor; fainter signal.

MTT note: ROI variance uses events, not hands. A winner might need 500–2,000+ tournaments for the same statistical comfort a cash player gets at 100k–250k hands (format-dependent; σ often 150–200+ in bb/100 equivalents).

Below ~10,000 hands, the normal approximation gets shaky; treat any CI as a rough sketch, not a contract.

Probability of a Losing Stretch While Winning

Even positive-EV players lose money over finite samples with surprising frequency. Approximate probability that cumulative result falls below zero:

$P(\text{loss}) \approx \Phi\left(\frac{-\text{WR} \times \sqrt{N/100}}{\sigma}\right)$

Where Φ is the standard normal CDF.

WR = +5 bb/100, σ = 90:

Hands	P(losing over sample)
20,000	~22% — more than 1-in-5
50,000	~11%
100,000	~4%

A 5 bb/100 winner losing over 50k hands is not cosmic punishment—it is a one-in-nine sample outcome. The math does not care that it feels impossible.

Expected profit vs swing band at WR = 5, σ = 90:

N	Expected profit	±1 SD band	2-SD worst case (bb)
50,000	+2,500 bb	±2,012 bb	−1,525 bb (~−15 BI)
100,000	+5,000 bb	±2,846 bb	−692 bb (~−7 BI)

These are distribution facts, not bankroll advice. They explain why two players with identical skill can post opposite graphs for months.

Monte Carlo Simulation

Closed-form CI assumes independent, stationary hands—poker violates that mildly (tilt, game selection, stake changes). Monte Carlo simulators (Primedope, PokerCharts variance calculators) bypass the algebra:

1. Input assumed WR, σ, and N
2. Draw thousands of random career paths where each 100-hand block varies normally around WR with width σ
3. Output percentile bands (70%, 95%), max downswing distributions, and risk-of-ruin estimates

What simulation adds that formulas hide:

Output	What it shows
Confidence cone	Fan of plausible bankroll paths—not one average line
Downswing depth percentiles	"95th percentile worst drawdown over N hands"
P(profit)	Empirical version of the Φ formula above
Sensitivity to σ	Same WR, σ = 75 vs 110 → wildly different cones

Run each scenario at σ_low and σ_high if your tracker sample is thin. σ is the most sensitive input in variance analysis.

Full Worked Audit

Profile: Online 6-max NLHE, N = 35,000 hands, observed WR = +3.8 bb/100, tracker σ = 95 bb/100.

Step 1 — Standard error

$\text{SE} = \frac{95}{\sqrt{350}} = \frac{95}{18.71} \approx 5.08 \text{ bb/100}$

Step 2 — 95% CI

$3.8 \pm 1.96 \times 5.08 = 3.8 \pm 10.0 \Rightarrow \textbf{−6.2 to +13.8 bb/100}$

Verdict: compatible with breakeven or with a strong winner. The sample is too short to upgrade stakes, too short to quit, too short to prove a leak does not exist.

Step 3 — Cumulative swing scale

Total 1-SD ≈ 95 × √350 ≈ 1,778 bb (~18 BI at $1/$2) around expected +1,330 bb profit. A −5 BI graph is inside the cone—not evidence of a broken strategy by itself.

Step 4 — Simulation homework

Plug WR = 3.8, σ = 95, N = 35,000 into a variance calculator; compare 95% cone width to your actual graph. If reality sits outside the simulated 95th percentile band, investigate σ mis-estimation, stake mixing, or non-stationary play before blaming variance.

What This Article Deliberately Omits

This is measurement science, not downswing operations:

• No stop-loss checklists, stake-move triggers, or tilt protocols—that is bankroll operations, not variance definition
• No "variance is an excuse for leaks" lecture—if VPIP/PFR drift >2 pts while graphs fall, the signal and the noise are both talking

When the CI finally excludes zero and graphs sit inside simulated cones, then operational articles about bankroll and psychology earn their place. First measure the noise; then decide whether strategy needs change.

Summary

Variance is the width of the error bar on your win-rate measurement. σ tells you how noisy your format is; N tells you how many hands you logged; CI tells you what win rates remain plausible.

Log hands. Pull σ from your tracker or bracket it with format defaults. Run the CI formula before you rewrite your identity as a winner or a loser. In poker, "enough data" is often 100k–500k hands, not a weekend heater.