[ Methodology ]
How Jonbar knows what it claims.
Jonbar reads price response and campaign lift from your sales history and reports every result as a range, locked before the outcome.
- Predicted range
- Outcome inside the range
- Outside
Starting point
Published research first, your history second.
Price response starts from published estimates and moves toward your own data as far as your past price changes allow. If your price never moved, the range stays wide and the brief says so.
Estimated effect of a discount on units →
- Range width ■■■■Published research only
- Range width ■■+ one season of your sales
- Range width ■+ two seasons of your sales
Mechanisms
Seven in the engine, three in the pilot.
The pilot uses price response, campaign lift and economics. The other four stay off until they can be calibrated, and the brief lists them as assumptions.
Select a mechanism to trace it.
Versions
Versions meet the same weeks.
Every version runs on the same random draws, so the gap comes from the decision, not from luck. The best-looking plan is pulled back toward the average, because the top of a search is partly noise.
Weeks 1Weeks 12
- Version A
- Version B
- Version C
Locked forecasts
Locked before the outcome.
The engine sees only what was known before one of your past campaigns, and the forecast is locked. Only then is the outcome opened and scored.
- 01Forecast lockedEngine version, inputs and seeds, with a timestamp.
- 02Outcome openedThe forecast can no longer move.
- 03ScoredLocked forecast against actual sales.
Details: published priors, limits, what the engine does not do
Three published priors
−2.62
average price elasticity across 1,851 published estimates.
~1/3
share of a promotion's extra units won from rival brands.
Hidden dip
the post-promotion dip is often invisible in store data. The pilot keeps it off.
Source: Hendel & Nevo 2003
Where the range gets wide
- A product with no history.
- A price that never moved.
- Few campaigns: lift is partly assumption.
- Thin data supports direction, not size.
- Shocks with no precedent in your data.
- Sales already summed across channels.
Not in the engine today
- Shopper-level simulation.
- Holiday and payday effects.
- Returns and ad cost.
- Learning across sellers.
- Measuring does not update the model by itself.
Language models only write the explanation. The best score 40.8/100 at simulating human behaviour.
Source: SimBench, ICLR 2026