How the model actually learns — in plain English Every time someone practices on GanttGrind and then takes the real PMP exam, they leave behind two things: a detailed map of how they prepared (every question answered, every score, how consistently they studied, which domains they drilled) and the outcome (pass or not). The prediction engine reads thousands of these maps side-by-side and learns which preparation fingerprints tend to end in passing. That's not a hand-wavy promise. It's a specific mathematical technique — logistic regression — that finds the relationship between your study patterns and the probability of passing. The model doesn't know the answer in advance. It discovers it from the data, one real exam result at a time. The blurry photo problem Here's the thing: not all practice profiles are equally informative. Imagine you're asked to identify someone from a photo. If the shot is taken from 200 feet away in bad light, you can make a rough guess — tall, dark hair, maybe. But you'd be wrong a lot. A close, clear photo is a different story. You can be certain. Practice data works the same way. A student who answered 50 questions before sitting the exam has a blurry profile. With that few answers, random luck plays a massive role in their accuracy — they might score 70% because they happened to draw easier questions that day, not because they're genuinely at that level. A student who answered 700 questions has a sharp, detailed portrait. Their patterns are real, not coincidental. The noise has averaged out. This is why the model doesn't treat all exam results equally when it trains. A rich practice profile — lots of questions, across many topics, over multiple sessions — carries more weight than a thin one. It's not about judging users. It's about how reliably the data reflects their true preparation. A profile with 50 questions contributes roughly one-tenth the influence of a profile with 500. Both are heard. One just speaks more quietly. The opinion poll problem Training a prediction model is like running a very large, very specific opinion poll. The more people you poll, the tighter your confidence in the result. But the composition of the poll matters just as much as its size. If a pollster only surveyed people who already planned to vote for one candidate, the results would be useless for predicting the real election. Our model faces an analogous risk: most serious GanttGrind users pass (~85%). If the model only ever saw passes, it would learn that "people who practice hard pass" — which is true but not useful. It couldn't tell you whether you specifically are ready, or just that prepared people generally pass. Non-pass results are what give the model its edges. They show exactly where the line is — not in some abstract theoretical sense, but in terms of real preparation patterns. The model learns: "Users with this readiness score and this domain coverage profile passed at 83%. Users with the same readiness but this specific weakness in Business Environment passed at 61%." That precision only comes from having the full picture. Every non-pass result someone shares is, in a very real way, a gift to every future candidate who's trying to gauge where they actually stand. Why Phase 1 predictions feel more like advice than numbers In the early phase — before the model has enough quality profiles — it does something different: it applies weights to your study patterns that reflect what we know about PMP preparation generally. Higher readiness matters more than question volume. Consistency matters more than single marathon sessions. It's informed reasoning, not statistics. Think of it like asking a seasoned instructor to assess you after looking at your study log. They can make a smart call based on experience and intuition. But they'd be more confident — and more specific — if they could also show you a database of a thousand students with similar logs and tell you exactly what percentage passed. That database is what we're building. What changes when the model has enough data At 30 quality profiles, something shifts. Instead of applying predetermined weights, the model starts discovering the actual relationship between features and outcomes from the data itself. Maybe it learns that domain coverage matters more than raw accuracy for a certain kind of learner. Maybe it learns that study consistency in the final two weeks is a better predictor than total questions answered. It doesn't know these things until it has enough data to see the patterns clearly. At 200 quality profiles, it can start cross-referencing. Not just "your readiness is 78% — here's what that band looks like" but "your readiness is 78%, your People domain is strong, and your Business Environment coverage is below 50% — here's what that specific combination looks like." Every additional dimension it can slice on makes the prediction more personal and more precise. One last thing: what "quality profile" actually means Not all uploads pair with a meaningful practice history. Some users upload their score report with very little GanttGrind practice behind it — maybe they were primarily using another resource and just wanted to log their result here. Those uploads are still valuable for post-hoc calibration (checking whether predictions were right), but they don't contribute much to training the model, because there's no detailed practice map to learn from. A quality profile is one where the person practiced enough questions — across enough different topics, over enough time — that their practice data actually reflects their real ability. That's the data the model learns from most. And that's why the best thing you can do for the whole community is to use GanttGrind seriously before your exam, and then upload your result when you're done — whatever it is. The model is a collective learning engine. The more people who contribute honestly, the more precisely it serves the next person who walks through the door.

Frequently Asked Questions

What is GanttGrind? Why "grind"?

Wait, ECO 2021 vs ECO 2026 — what's the deal?

Which version should I practice with?

How are your questions made?

Found a question error — what do I do?

How does "adaptive" learning actually work?

What's the readiness score?

What exam sections do you cover?

What type of questions are these?

Should I upload my score report (even if I didn't pass)?

How do I upload my score report?

How does the "likelihood to pass" prediction work?

How does the prediction get smarter over time?

Can I use this like an app on my phone?

Is this free?

What's your refund policy?

What if I can't afford premium?

I want a feature that doesn't exist. Can you build it?

How do I get in touch?