Predicting sales for a product that doesn't exist yet is one of the hardest problems in data science, and the current approach, training an XGBoost regressor on historical invoices, deserves the team's skepticism. Without actual sales data for the new innovation, any model built on past patterns is essentially guessing which features will generalize to an unseen context. This isn't a failure of the tool; it's a failure of the framing.
The challenge here isn't technical, it's conceptual. An XGBoost model trained on month-level invoices, weather, and macroeconomic data can capture correlations in what *has* happened, but it cannot learn what *will* happen for a product with no history. The team is right to lack confidence. The real question is whether they can shift from a predictive paradigm to a generative one. Instead of asking "What will sales be?" they should ask "Given these attributes and conditions, what range of outcomes is plausible?" This is where probabilistic modeling or scenario-based simulation becomes more honest than a single point estimate. The approach used by Training an AI on 100 million square feet of building data to transform construction offers a useful parallel: instead of treating past data as a direct predictor, that system builds a world model grounded in physics and standards, allowing it to simulate outcomes for designs that have never been built. A similar logic applies here, construct a simulation layer that encodes product attributes, regional conditions, and macroeconomic levers, then run thousands of scenarios to map the probability distribution of sales.
The temptation is to add more data or tune hyperparameters, but that treats the symptom, not the cause. A more productive path is to treat the launch as an experiment. Use the XGBoost output as one input into a Bayesian framework that incorporates prior beliefs from category analogs or expert judgment. This makes uncertainty explicit rather than hiding it behind a regression score. The team can then say, "Based on this model, there's a 70% chance sales fall between X and Y," which is far more actionable than a single number no one believes. Waymo and Toyota partner to test autonomous taxis in Japan demonstrates this principle: they are testing in a controlled environment before scaling, building confidence through iteration rather than prediction. The product launch team should adopt the same mindset, treat early months as a learning phase, update the model with real data as it comes in, and resist the urge to commit to a forecast before the market speaks.
The specific takeaway is this: stop trying to predict the unpredictable with a point estimate. Build a simulation that generates a range of plausible outcomes, update it with real data post-launch, and communicate the uncertainty directly to stakeholders. That is a smarter, more honest approach than clinging to a model no one trusts.