future product-week-price cells compared against empirical 2023 sales
Pricing & Promotion: East-Asia FMCG
A year-ahead pricing model, validated against real sales.
A global FMCG manufacturer withheld 2023 weekly unit sales for four priority products as a blind out-of-time holdout. Simulacra trained on 2021–22 NielsenIQ pricing and promotion data alone, predicted every observed 2023 product-week-price cell, and the customer scored the predictions against the empirical holdout. Year-ahead unit scale, price-response direction and magnitude, and portfolio bias all replicated.
median absolute weekly unit error across the holdout
of predictions inside each product's own weekly sales volatility
mean signed unit error across all cells, effectively centered at zero
Verified
Customer and SKU names anonymized at the customer's request. Every plotted row, prediction, price, and error metric on this page is an exact value from the 2026 customer-reviewed validation output.
Protocol: out-of-time holdout design
Train on 2021–22 → Predict 2023 → Compare to empirical sales.
Blind time-delayed holdout
Simulacra saw weekly sales, price, discount, and market-price context for 2021–2022. Empirical 2023 units were held out by the customer and used only for final comparison.
250 synthetic rows per product-week.
For each observed 2023 price cell, Simulacra generated the implied weekly demand distribution and returned the predicted unit volume.
Score locked predictions against the unsealed holdout.
Simulacra generated predictions blind. The customer revealed 2023 actuals, validated the predictions, and signed off on these results.
Validation 1: year-ahead unit scale
Predicted weekly units track held-out 2023 sales across four products.
Each point below is one future product-week-price cell. The diagonal is perfect prediction; the shaded band is that product's empirical weekly standard deviation. Simulacra lands 83.1% of all predictions inside that business-relevant volatility band, with a median miss of 172 units.
Predicted vs empirical weekly unit sales, 2023 holdout
195 cells, four customer-selected products, y = x is perfect prediction, SKU identities blinded
Validation 2: price-response shape
Simulacra learned the full demand curve.
Pricing research is only useful if it gets directional economics right. Across all four products, Simulacra predicted the correct negative price-volume slope. The predicted slope magnitude landed between 69% and 110% of the empirical 2023 slope.
Weekly volume vs price per unit, empirical vs Simulacra
195 cells; four products; teal = empirical 2023, magenta = Simulacra prediction; slope-match labels show predicted slope as % of empirical
Validation 3: calibration profile
Near-zero portfolio bias, with error calibrated to weekly volatility.
A model can look good on average while leaning high or low. This one does not: mean signed error across all held-out cells is +9 units. The residual distribution is centered near zero, and 82.6% of product-week predictions land within ±500 units.
Prediction error distribution, predicted − empirical weekly units
Shaded bands show ±200 and ±500 weekly units. Largest miss annotated as the typhoon week.
Known limitations
External events drive largest future-prediction discrepancies.
The validation includes the typhoon week, the under-recovered slopes, and every cell outside the weekly-volatility band. These are the gaps a buyer should see before trusting the result.
−1,695 units on the September 2023 typhoon week.
The largest single miss in the holdout, on the highest-volume product. The 2021–22 training window contains no comparable external shock, so Simulacra could not anticipate the disruption. The miss is annotated on the residual histogram and is not excluded from any aggregate metric reported on this page.
Direction right. Magnitude under-recovered by ~30%.
Price-response sign was correct for all four products. The predicted slope magnitude on Products 1 and 3 landed at 69% of the empirical 2023 slope — Simulacra captured the price sensitivity directionally but understated how steep it was. Products 2 and 4 fit tighter.
33 of 195 cells fell outside their product's weekly volatility band.
The remaining 16.9% of cells missed the ±1-SD threshold reported in Validation 1. The signed errors across these misses remain centered near zero, so the residuals are noisy rather than directional — but they are real, and they include the typhoon week.
Generative Causal AI vs conjoint
The next generation of market simulation is scored on behavior, not just stated choice.
Conjoint and pricing simulators help teams reason about willingness to pay, share, revenue, and demand curves. Simulacra pushes that workflow further: learn the response structure from real commercial or research data, run pricing scenarios against it, and validate the predictions against an empirical holdout.
Survey trade-offs.
Great for new products and unobserved attributes; still depends on stated choices and simulator assumptions.
Response surfaces from data.
Learns how price, promotion context, product, and demand move together, then generates counterfactual product-week outcomes.
Year-ahead holdout.
This validation scores a future-year weekly POS holdout against the customer's empirical sales data.
Validate on your data
Bring two years of weekly POS. We'll predict last year's sales — and find the changepoint where Simulacra starts paying for itself.
Bring weekly POS, pricing-and-promotion history, a tracker, conjoint, or U&A. You choose the hidden slice. We fit on the rest, deliver predictions, and compare them to the empirical holdout.
About the customer:
East-Asia FMCG manufacturer, NielsenIQ Pricing & Promotion data, four-product snacks portfolio, name under NDA.
Validation outcome: year-ahead weekly demand prediction across 195 held-out cells. Customer-reviewed; all figures reproducible from the 2026 blind validation output.