How It Works

Monte Carlo simulation

Some probabilities you can derive with algebra. Plenty you can’t, because the math turns brutal or there’s no closed-form answer at all. So instead of solving the problem exactly, you let the scenario play out thousands of times with fair random inputs and count how often each outcome happens. The fraction of runs where a thing occurred is its probability, to within a small and shrinking margin of error.

That’s Monte Carlo simulation, and it’s the engine behind the lab’s analytics work. TankOdds runs the NBA draft lottery and the rest of the schedule tens of thousands of times. Every published number is a count over those runs, not a formula. When the standings change, the whole distribution moves with them.

Distributions, not point estimates

A single number hides how uncertain it is. “Projected to finish 4th” tells you nothing about whether 1st through 7th are all live. The lab reports the full spread instead: “22% chance at the 4th pick, with everything from 1 to 7 in play.” That’s the shape you need to make a decision. The error in a simulation shrinks with the square root of the number of trials, so a rough answer is cheap and a precise one costs real compute. The lab pays that cost so the published figure is one you can lean on.

Biostatistics you can defend

For medical-education outcomes, the lab uses conventional inferential statistics rather than simulation, because the exact tests exist and a grant reviewer expects them. CMEye picks the right test for the data it’s handed and reports effect sizes next to p-values, never a bare p-value dressed up as an answer. That behavior is the default because the person who built it has had to defend statistics to a reviewer.

Reference and evidence discipline

Every citation in a CME proposal is NLM-formatted and full-text verified, then audited for industry-sponsor bias. If a claim can’t be verified on its own, it gets cut. The same rule covers the numbers on this site. Nothing goes up that the lab can’t stand behind.

The honest limitations

A model is only as good as its assumptions. Get them wrong, say the sample doesn’t represent the population or the events aren’t independent, and the simulation still converges, beautifully, on the wrong answer. Precision and accuracy aren’t the same thing. Every model the lab ships states its assumptions so you can judge them yourself.

Questions about any of this? The FAQ goes further, or you can just ask.