By Persis Duaik, Presales Consultant at Panintelligence and the Brazilian half of our World Cup prediction team
In the previous blog, I explained how we prepared more than 49,000 historical international football results for analysis, turning raw facts into model-ready variables: form difference, attack and defence ratings, and tournament importance. We even built the Panintelligence Team Strength Index (PI-TSI) to give us a clean, transparent baseline. But let's be honest: creating a model is only half the battle. A spreadsheet with thousands of fixtures might make Reece happy, but it is not going to inspire anyone else to explore the data. We needed to bring it to life.
Making the decision tree visible
That is why we presented the analytical model through a Sankey-style decision tree visualization. Instead of just spitting out a static prediction, the chart lets you trace the model's exact reasoning. You can see which variables trigger each split, the thresholds applied, how many historical matches went down each path, and where the patterns are strong versus where uncertainty takes over.
The goal is simple: if the model makes a prediction, users should be able to see how it got there. A traditional prediction just hands you a percentage. The decision tree lets you look under the hood and ask: Why does the model think that? For now, we are keeping the exact World Cup paths under wraps, but the logic is out there in the open not hidden in some mysterious black box.
Turning the analysis into team strength
The next step was the interactive Team Strength view. Every single participating country gets a PI-TSI score so you can see the entire field at a glance. But we did not want a flat, static ranking. In our dashboard, you can select any country (obviously, I went straight to Brazil and Morocco) and inspect the underlying Form, Attack, and Defence indicators that make up their scores.
This gives you two levels of analysis: compare the overall standings and then drill down into the performance factors. It is the difference between publishing a set-in-stone verdict and inviting a real debate. Rather than presenting a final verdict, we wanted users to be able to challenge the conclusion for themselves.
Simulating the group stage
Simulating the group stage turned out to be more complicated than predicting the matches themselves. Of course, the World Cup is not just a list of teams it is a 12-group tournament. To simulate it, we had to translate predicted match outcomes into group standings: three points for a win, one for a draw, and zero for a loss. But things get complicated when teams tie. If points are level, the simulator runs through our tie-breakers: head-to-head points, goal difference, goals scored, and finally the PI-TSI score itself. The top two teams from each group qualify automatically, alongside the eight best third-placed teams.
This was a healthy reminder that predicting individual matches is only one piece of the puzzle. Every result kicks off a chain reaction: the match outcome shapes the points, the points determine the group rank, and the group rank decides who advances to the knockouts. In business, predictions behave the same way. A customer risk score does not exist in a vacuum; it triggers an automated alert for the success team to reach out. The analysis is only as useful as the process it feeds into.
Building the knockout bracket
The knockout allocation rules were one of the first places where a working model met real-world complexity. Once the group stage wraps, the real tournament begins: the knockout bracket. We mapped the qualifying teams directly into the official FIFA match-number pathways from the Round of 32 all the way to the final. It would have been far easier to just draw a pretty, static bracket and hope nobody checked our math, but here at Panintelligence, this is not an option. The simulation handles the complex combinations of qualifying third-place teams, mapping them to their correct Round of 32 slots and simulating each high-stakes match. This creates a fully connected prediction journey from matchday one to the final whistle.
The refinements we made
Moving from a Python model that runs on a laptop to a production-grade interactive dashboard exposed plenty of real-world messiness. Here is how we tackled it:
- Standardizing the Data: Historical datasets are messy. Teams change names or use different spellings depending on who recorded them. We used a mapping dictionary (TEAM_ALIASES) to standardize entries like "USA" to "United States", "Congo DR" to "DR Congo", and "Korea Republic" to "South Korea" so the database, charts, and frontend always talk to each other cleanly.
- Managing the Small-Sample Problem: If a team has only played one or two recent games (for instance, Cape Verde or Uzbekistan), their recent stats can look artificially elite or disastrous. We introduced a data_confidence multiplier, calculating it as LEAST(recent_match_count / 5.0, 1). If a team has fewer than five recent matches, their Form, Attack, and Defence scores are scaled back toward a neutral baseline of 50. This stops short-term noise from skewing the rankings.
- Sizing the Performance Window: For calculating both Form and Attack/Defence metrics, we settled on a rolling window of the last 5 international fixtures. It is recent enough to capture momentum, but long enough to filter out single-match flukes.
- Handling Extra Time and Penalties: In our historical match analysis, we look at the goals scored at the end of play (including extra time, but excluding penalty shootouts). For the knockout stages, matches must have a winner. To keep it clean and explainable, our simulation determines the winner deterministically based on the PI-TSI score difference, with the higher-rated team advancing with a simulated 2-1 scoreline.
- Accounting for Close Matchups: If two teams are closely matched in the group stage, predicting a clear winner is unrealistic. We built a 5-point buffer: if the difference between their PI-TSI scores is less than 5, the simulator predicts a 1-1 draw.
- Navigating the 3rd-Place Maze: Qualifying the best eight third-placed teams across 12 groups was a logic puzzle. We ranked all third-place finishers by points, goal difference, goals scored, and PI-TSI. Once the final eight were selected, we mapped them into the official Round of 32 matchups using the complex FIFA allocation table.
- Smoothing the User Experience: Building a 32-team bracket in HTML/CSS that fits on a screen is tough. We split the dashboard into clean tabs to avoid cognitive overload, added a dynamic dot-navigation sidebar that highlights your location as you scroll, and built loading states with placeholders so users are not left staring at blank spaces while the iframes fetch live charts from the Panintelligence server.
- Refining the Weights After Review: When Reece and I ran the feature importance analysis on our historical database, we adjusted the PI-TSI weights. We shifted from our initial guess of 53% Form, 31% Defence, and 16% Attack to the actual data-driven coefficients: 52% Form, 30% Defence, and 18% Attack.
Human judgement still matters
A model finds patterns, but it does not understand context. That requires human review. Reece and I challenged the assumptions, tested the simulation edge cases, and made sure every single number could be traced back to real evidence. That does not mean changing a prediction because we do not like it sorry Reece, the model does not care about your England bias. But it does mean ensuring the logic matches the tournament's spirit. Predictive analytics should guide human judgement, not replace it. (Though if Brazil and the model disagree in the knockout stages, I reserve the right to trust my heart!)
The continuing story
The most exciting part of this project is not our pre-tournament prediction; it is what happens when our model meets reality. The model will make its call, the players will play, and every result will feed fresh data back into the loop. We will update the standings, recalculate the PI-TSI, and see how the predictions shift.
So, what is it actually predicting? In the next blog, we will reveal how the model performed during testing, which teams it rates the highest, the potential group-stage surprises, and the simulated routes to the final for both England and Brazil. The dashboards are live, the code is set, and Reece is still checking the numbers. I am busy checking Brazil's path to the trophy. Purely for QA, of course.
