Describe a strategy for reconciling different forecasts into one robust prediction

WHAT THIS TESTS: The interviewer wants to see if you treat forecasting as a meta-learning problem rather than a model-selection problem. They are looking for sophistication in combining heterogeneous signals, reasoning about uncertainty, and communicating probabilistic outcomes to leadership. The core skill is synthesizing top-down market logic with bottom-up operational mechanics without collapsing into a false sense of precision.

A GOOD ANSWER COVERS: First, diagnose the divergence. A senior candidate explains that top-down and bottom-up models often differ because they rely on independent assumptions such as market growth rates versus cohort retention curves, and the reconciliation should start by mapping where the gap originates. Second, choose a weighting scheme. Options include inverse-variance weighting if you have historical forecast errors, Bayesian model averaging if you can define model priors, or simple stacking where a meta-model learns optimal weights from past performance. Third, model the dependency structure. If both models rely on the same macro variable they are not independent, so a straight average double-counts that risk; instead, use a covariance-aware ensemble or mixture distribution. Fourth, output a distribution, not a point. Present leadership with a fan chart or prediction interval that reflects model disagreement as epistemic uncertainty rather than hiding it behind a single number.

COMMON WRONG ANSWERS: A major red flag is proposing a simple arithmetic average without discussing why the models differ or how correlated their errors are. Another weak pattern is picking the model that feels more conservative without quantitative justification. Some candidates suggest building a third complex model to break the tie, which misses the point that ensemble methods are explicitly designed to leverage existing forecasts efficiently.

LIKELY FOLLOW-UPS: The interviewer might ask how you would update weights as new actuals arrive, which should trigger a discussion of online learning or exponential decay of older errors. They might also probe what you do when models have non-overlapping confidence intervals, which is a signal to audit assumptions rather than mechanically blend. A third follow-up could be how you present a wide consensus range to executives who want a single target, which tests your ability to translate uncertainty into decision frameworks like scenario planning.

ONE CONCRETE EXAMPLE: Suppose your top-down model forecasts ten million dollars in revenue based on a two percent market share assumption, while your bottom-up cohort model forecasts seven million dollars based on flattening activation rates. Instead of splitting the difference at eight point five million, you examine the sensitivity of each model to its key driver. You discover the top-down model has historically overestimated by fifteen percent due to optimistic market sizing, while the bottom-up model is unbiased but noisier. You apply an inverse-variance weight that gives sixty percent weight to the bottom-up forecast and forty percent to the top-down, then you run a Monte Carlo simulation that samples both driver distributions to produce a consensus forecast of seven point eight million dollars with an eighty percent prediction interval from six point five to nine point two million dollars.