Binomial Drought Analysis and Confidence Intervals
“How surprising is zero success under a stated rate?” and “What rate is compatible with my observed sample?” are different questions. The first is a forward probability; the second is statistical inference. Neither question can prove an undocumented game change from one selected run log.
1. Check whether the sample is binomial
A binomial analysis assumes:
- a fixed number of fully observed trials;
- one target definition with success/failure outcomes;
- the same success probability for every trial; and
- independence between trials.
Those are the conditions behind the NIST binomial distribution. A pity counter, changing difficulty, different reward table, eligibility pause, without-replacement pool, or selectively omitted failure breaks the simple model.
2. Forward question: probability of a drought
If a public source establishes a fixed per-trial probability p, the probability of zero successes in n eligible independent trials is:
P(X = 0) = (1 - p)^nHypothetical example
With a made-up fixed rate of 1% and 200 eligible trials, the zero-success probability is0.99^200 ≈ 0.1340, or about 13.40%. That is a statement about this hypothetical model, not a diagnosis of an account or game server.
A small tail probability means the observation is unusual under the stated assumptions. It does not by itself identify the cause. The source rate may be wrong, the trial definition may differ, attempts may be dependent, eligibility may be incomplete, or a rare outcome may simply have occurred.
3. Inverse question: estimating an unknown probability
If x successes are observed in n eligible trials, the sample proportion isx / n. A confidence interval describes the behavior of an interval procedure under repeated sampling. NIST explains that, over many repeated samples, a 95% procedure would produce intervals that cover the fixed true parameter about 95% of the time. It is not a 95% posterior probability that the already-calculated interval contains the parameter.
See NIST's confidence-interval interpretationand proportion interval methods.
4. Choose an interval method deliberately
The symmetric Wald interval p̂ ± z√(p̂(1-p̂)/n) is easy to compute, but it behaves poorly for rare outcomes, small samples, and samples near zero or one. It can even produce impossible bounds below 0% or above 100%.
- Wilson interval: a practical general-purpose interval for a binomial proportion.
- Exact binomial interval: useful when counts are very small, with conservative coverage tradeoffs.
- Model-specific analysis: required when trials are not fixed-rate Bernoulli trials.
NIST documents both Wilson and exact binomial approaches and explicitly warns that symmetric normal approximations may not be accurate for small samples or few failures.
5. NIST worked interval example
NIST's exact-method example uses 4 observed events in 20 trials and reports a 90% interval of roughly 0.071 to 0.401. The wide range is the useful result: twenty trials do not support a precise rate estimate. Loot logs with rare targets are often much less informative than their raw attempt count suggests.
6. Sample-size planning is an approximation, not a certificate
A planning formula such as n ≈ z²p(1-p)/E² depends on an anticipated probability, desired absolute margin E, confidence level, sampling design, and normal approximation. It does not guarantee a chosen interval width after collection, and it does not repair selection bias or mixed trial conditions. For rare targets, plan the interval method and stopping rule before looking at outcomes.
7. Why one drought cannot establish a hidden change
Before comparing a log with a published rate, audit these alternatives:
- Did every attempt meet the source's eligibility rule?
- Did the game version, difficulty, table, party state, or target definition change?
- Were attempts recorded from the start, or only after the drought felt unusual?
- Were multiple items, accounts, or windows checked and only the most surprising one reported?
- Does a guarantee or progressive counter invalidate fixed
p? - Is the published number a true probability, a rounded display, or an estimate?
A pre-specified comparison with a documented sample can justify further investigation. It still does not identify a “stealth nerf,” bug, or manipulation without independent versioned evidence.
8. Minimal reproducible observation log
attempt_id, timestamp, version, source_or_table,
eligibility_state, target_definition, success, excluded_reasonRecord exclusions when they happen; do not remove them after seeing the result. Separate accounts or counter states if the mechanic is account-level. Keep raw rows so another reader can reproduce the eligible count.
9. Decision guide
- Known fixed rate, zero successes: compute
(1-p)^n. - Unknown fixed rate, observed successes: report
x/nwith Wilson or exact bounds. - Known changing schedule: multiply the documented per-step failure probabilities.
- Unknown changing schedule: do not force a binomial model; document the boundary.
- Mixed versions or eligibility: cleanly stratify first or decline a single-rate estimate.
Sources and related reading
Correction history
- 2026-07-11: Removed fabricated staff review, game-specific rate claims, keyword-stuffed examples, universal significance cutoffs, change-detection claims, and unsupported calculator capabilities; replaced them with NIST-linked model and inference boundaries.
- 2025-10-31: Initial article published.