Reading the Confidence Band

Bob had opened the Scout tab on three players in his stable. Three bell curves, three wildly different widths. He wanted a rule for what to do with each. Tau drew him a triage chart on a napkin.

Same booth. The waiter had stopped trying entirely and was just wiping glasses at the bar. Tau had a pen in one hand and a horchata in the other.

Bob: OK I'm on the Scout page. Got three guys pulled up. One curve looks like a thumbtack, one looks like a normal bell curve, one looks like a blob that's been hit by a truck. What do I do with three different shapes?

Uncle Tau: Three different things. The shape is the decision. The centre is a suggestion.

Bob: Everybody stares at the centre.

Uncle Tau: Everybody is wrong. The centre is the point estimate — the app's best guess. The width is the app telling you how much to trust the best guess. Narrow band, trust it. Wide band, don't. In between, think harder. That's the whole lesson.

Bob: You're about to give me three buckets, aren't you.

Uncle Tau: I am.

The three triage buckets

Uncle Tau: Every posterior you see in this app — Results, Strategy, Scout, Package Builder, Stable — falls into one of three bands. Call them act on, directional, and come-back-later.

Act on. The band is tight. The centre is informative. You can trust the number and build a decision around it. Stake the guy, adjust your markup, size your action pool, back your teammate. The app is quoting you a player who has enough data to speak for themselves.

Directional. The band is moderate. The centre is a defensible guess but not a committed one. You can use the sign and the rough magnitude but you should price in the uncertainty. Stake smaller than the centre suggests. Take a conservative markup. Size the action pool for the possibility that the true number is meaningfully different.

Come-back-later. The band is a barn door. The centre is barely more informative than the prior. Don't build decisions off this. Play more, wait more, or explicitly use the prior and acknowledge you're doing that. The app is telling you it doesn't know.

Bob: And I can see all three on one page.

Uncle Tau: You often will. Three players, same format, wildly different $n$. One curve a thumbtack, one a bell, one a blob. Each one calls for a different kind of action.

Why the band shrinks at the rate it does

Bob: Why is the band width even a real measurement? Why not just "the app is more confident now."

Uncle Tau: Because it's not a preference. It's a measurement floor. The band width is the posterior standard deviation, and the posterior standard deviation has a hard floor — Cramér-Rao. It can't shrink faster than $1/\sqrt{n \cdot \mathscr{I}}$ where $\mathscr{I}$ is the Fisher information per observation. Every unbiased estimator is stuck behind that floor. The app's band width is basically sitting on that floor.

Bob: So the band is only as tight as the math allows.

Uncle Tau: And not tighter. If a different tool shows you a tighter band on the same sample, it's either using a tighter prior — leaning on assumptions — or it's lying. Those are the only two options. The app's band is Cramér-Rao honest. Narrow priors get you tighter bands but only by committing to assumptions that your data can't refute yet.

Bob: And the math says the band shrinks like one over the square root of n.

Uncle Tau: For most of the parameters you care about. Which means if you want to halve your band, you need four times the sample. To quarter it, sixteen times. The marginal cost of precision grows fast. At some point the band is wide enough that more data is the only way forward, and at another point the band is tight enough that more data barely moves the needle.

Bob: So I'm always inside that curve.

Uncle Tau: Forever. Every player. Every format. Every parameter. There is no escaping the $\sqrt{n}$ rate. The only choice is which bucket you're currently in, and whether to act on it.

Where you see this across the app

Bob: OK so "bucket your decisions by band width." Where does that actually show up on the pages I look at.

Uncle Tau: Everywhere you get a number. Let me walk you through.

Results. Your personal posterior. The band around your ROI curve is the width bucket for you. If your band is thumbtack-tight, you can quote your own number confidently. If it's a blob, your real ROI could be materially different from what the centre says, and bragging about the mean is just noise laundering.

Strategy. When SALSA runs, it doesn't pick one ROI and simulate. It samples from your posterior a few thousand times and simulates each. Every output range you see — growth rate, Kelly fraction, optimal selling fraction — is wider when your posterior is wider. If the strategy output looks indecisive, it's because your posterior is indecisive. Narrow the input, narrow the output.

Scout. Someone else's posterior. Three thousand tournaments on them — tight band, act on. Sixty tournaments on them — barn door, come back later. The app isn't being lazy about new players. It's being honest.

Package Builder. Uses posterior width directly to price markup bands. Narrow posterior, tight markup window — the seller and the buyer are negotiating over a parameter both sides can see. Wide posterior, wider markup band — because both sides are negotiating over a belief that neither side can pin down.

Stable. Per-teammate posteriors with the same three buckets. Some teammates are in act-on territory for your pool sizing. Some are directional. Some — the new ones — the app is essentially showing you the prior plus a nudge, and you should stake accordingly.

The honest reading

Bob: When people complain the app "doesn't give me a clear answer" — they're complaining about a wide band.

Uncle Tau: And the wide band is the app refusing to make up a clear answer it doesn't have. Every other tool in this industry gives you a clear answer at small $n$. Graph of 40 tournaments, single number on top. That's not clarity. That's confidence theatre.

Bob: I've quoted point estimates with a straight face plenty of times.

Uncle Tau: Everybody has. Everybody's a hypocrite. The app stops you from lying to yourself by drawing the width. If the band is wide, that's information. Pretending it isn't wide doesn't make it narrower. It just transfers the uncertainty from the estimate to your behaviour — you act decisively on a number that doesn't support decisive action, and variance does the rest.

Bob: And that's what "honest admission of ignorance" means.

Uncle Tau: The posterior width is the app's ignorance admission, written down in buy-in units. If you read only the centre, you're reading the app with its mouth covered. The width is the rest of the sentence.

What about the Delta Method

Bob: You mentioned the band width connects to something called the Delta Method.

Uncle Tau: Downstream derived quantities. The app doesn't estimate ROI directly — it estimates cash frequency and HU winrate, and ROI is a function of those two. So the posterior uncertainty on ROI is a function of the posterior uncertainty on the two underlying parameters. The Delta Method is how you propagate that uncertainty through a smooth transformation.

Bob: Translate.

Uncle Tau: If $\mu$ is a smooth function of $(\lambda_1, \lambda_2)$ — and it is, through the SALSA partition function — then the posterior variance on $\mu$ is:

$$\text{Var}(\hat{\mu}) \geq \frac{1}{n} \cdot \mathbf{J}^\top \mathscr{I}(\lambda)^{-1} \mathbf{J}$$

where $\mathbf{J}$ is the gradient of $\mu$ with respect to the parameters and $\mathscr{I}(\lambda)^{-1}$ is the inverse Fisher information matrix. The inequality is tight as $n \to \infty$. Practically: the app computes this and draws the band.

Bob: So the band on my ROI is a compound object that inherits the tightness of the tightest parameter and the looseness of the loosest.

Uncle Tau: That's exactly what it is. Cash frequency converges fast. HU winrate converges slow. The ROI band inherits both. For most of your career the ROI band will be dominated by how little the app knows about your HU winrate. That's why even a 2,000-tournament sample leaves a posterior that isn't yet a thumbtack. You're feeling the HU winrate bottleneck.

Bob: And the only way to narrow it further is volume.

Uncle Tau: Volume and time. Cramér-Rao does not negotiate. The floor is the floor.

A working decision rule

Bob: Give me a simple one-line rule.

Uncle Tau: Look at the band first. If the band doesn't cross zero and isn't huge, act on the centre. If the band crosses zero or is wide, treat the number as directional and size down. If the band is basically the width of the prior, you're looking at the prior — treat that number as a best-guess assumption, not a finding.

Bob: And for staking?

Uncle Tau: Same rule. Tight band on the player, you can stake at close to the posterior mean. Wide band, you're buying uncertainty — the price should reflect that. Blob band, you're basically staking on the prior plus vibes — which you can do, as long as you know that's what you're doing and you price accordingly.

Bob: And for my own ROI —

Uncle Tau: Quote the band, not the centre. Anybody who asks you "what's your ROI" and gets a single number instead of a range is being lied to. Honest answer is a range. The range comes from the app. The range is your posterior.

Bob: So the bands on every screen are the same object, it's just the parameter that changes.

Uncle Tau: Same object. Same math. Different parameter. Your ROI. A teammate's cash frequency. A scouting target's Kelly fraction. A package markup band. All of it is posterior-variance plumbing. The width is the app being honest about what it knows. Read the width first, always.

Bob: The width is the ignorance admission.

Uncle Tau: And the ignorance admission is what makes the centre trustworthy. If the app wasn't willing to draw wide bands when it didn't know, you couldn't trust the narrow bands either. Consistency of honesty. Same reason a witness who admits when they didn't see something is more credible than one who claims to have seen everything.

Bob: Got it. Thanks, Uncle Tau.

Uncle Tau: Go estimate your shapes, kid. And next time, before you stake someone, look at the band before you look at the centre. The centre is the pitch. The band is the reality check.

What's next

Kelly under uncertainty — when the band is wide, the growth-rate-optimal thing to do is size smaller. Jensen's inequality says so. Here's how the app bakes that into every sizing output.
Game selection with a posterior — why you should prefer formats where your band is already narrow over formats where it's still a blob, and when to make exceptions.
Selling action to rebalance — the MOTA selling-fraction optimum is itself a function of your posterior width. Wider band, sell more.