Model the people in the survey
Around one thousand synthetic people were distributed across five cities and six socioeconomic levels.
Public case study · Colombia · 2025–2026
Could a simulated market anticipate a real one?
We used Claviata to predict Colombia's national consumer-confidence survey before the results were available. This case shows the prediction beside reality, plus the earlier approaches that failed.
This experiment was originally published under the name Calibra. Claviata continues that same line of work.
~1,000synthetic people per month
7past months used to test the method
5cities represented
2predictions made before publication
Prediction versus reality
We predicted a sharp rise. Then the survey confirmed it.
Before the June 2026 survey was published, Claviata predicted a score of +21.5. The survey later reported +24.3.
2.8points apart
Claviata predicted 21.5. The survey reported 24.3. The closer these numbers are, the better the prediction.
Method
Survey result: +24.3
How far off
Unadjusted AI panel
-15.6
39.9
Average of past results
+17.3
7.0
Same as prior month
+17.8
6.5
Claviata method
+21.5
2.8
One accurate prediction does not prove the method will always be right. It does show a real, timestamped result that anyone can inspect.
How the test worked
Four steps kept the test honest.
Limit what the system could know
Each simulation received only information that was publicly available during that month.
Record the answer before publication
The prediction was timestamped before the real survey result appeared.
Compare the numbers
Every prediction was placed beside the result published by Fedesarrollo and simple alternatives such as repeating the prior month.
What failed
The first approach failed. We published that too.
Before adjustment, the panel was 81.4 points away from reality on average. Simply repeating the previous month's result missed by only 3.2.
The synthetic people reacted too strongly to alarming headlines. The original score was not usable as a forecast, even though the differences between groups still contained useful information.
81.4 points
average miss before adjustment
The headline score failed.
0.45
movement with the real survey
Some patterns moved together.
67%
correct ordering between groups
Relative preferences were more useful.
What survived
What remained useful: who disagreed, and why.
After adjusting the method using past results, its average miss on held-out months fell to 1.44 points.
The biggest value was not one market-wide average. It was the consistent difference between income levels, cities, and response types. Claviata focuses on those differences so teams can decide whom to target and what to test.
What Claviata delivers
Turn one decision into an action-ready market read.
Describe the decision and choose the audience. Claviata shows how different segments of synthetic people respond, what persuades them, what stops them, and which assumption you should test next with real people.
- An interest score explained in plain language
- The segments most and least likely to respond
- Objections and adoption reasons in segment language
- A specific next step for a real-world test
Honest limits
What this study does not prove.
This study covers one country, one survey, and a limited number of months. One accurate prediction may still involve luck, and this evidence does not mean every market decision can be forecast with the same accuracy. Use Claviata to guide the next test with real people, not replace it.
