Predictive Forecasting & Simulation

March 2020. Every forecast model at eBay Classifieds broke overnight. Real estate listings collapsed. Auto demand evaporated. The historical patterns every model relied on became irrelevant in a week.

That is when you learn the difference between a forecast and a plan. A forecast tells you what will probably happen. A plan tells you what to do when it does not.

Why most forecasting projects fail

Companies invest in forecasting when they are tired of being surprised. Fair. But most implementations fail for the same reason: they build a single-point prediction ("Q3 revenue will be $12.4M") instead of a scenario framework ("Here are the three most likely outcomes, the conditions that trigger each, and what we do in each case").

Single-point forecasts create a false sense of precision. They also become political — whoever owns the model owns the number, and nobody wants to be wrong.

What we build instead

We build models that serve decision-makers, not data scientists. That means:

Scenario simulation, not point estimates. For eBay's five emerging markets during COVID, we built "no-COVID" baseline patches — synthetic historical patterns stripped of pandemic effects — then layered recovery scenarios on top. CFOs and GMs used these weekly to adjust budgets in real time, not quarterly.

Domain-specific algorithms, not off-the-shelf. At P&G, Walmart's automated replenishment system was creating a negative feedback loop: when products sat in the backend instead of shelves, sales dropped, the system read "low demand" and ordered less, creating a perpetual decline. We built a custom On-Shelf Availability algorithm that distinguished real low demand from availability gaps. The fix — strategic overstocking of 20 stores — generated $3M in incremental revenue in 4 months.

Monthly model reruns, not one-time delivery. A forecast model that is not retrained on fresh data degrades fast. Our churn model at eBay ran monthly against updated listing, traffic, and quality data — each cycle sharpening the retention targeting that delivered a 15% lift.

The methods behind it

We work in Python and R depending on the use case. Time series models (ARIMA, Prophet, custom exponential smoothing) for demand and revenue. Classification models (gradient boosting, logistic regression) for churn and risk scoring. Simulation frameworks (Monte Carlo, scenario trees) for planning under uncertainty.

The tooling matters less than the framing. Before we write a line of code, we define: what decision does this model serve, who will act on it, how often does it need to refresh, and what is the cost of being wrong by 5% versus 20%?

What you receive

A production-ready model with clear inputs, outputs, and retraining cadence. A scenario dashboard where non-technical stakeholders can adjust assumptions and see projected outcomes. Documentation that explains what the model does and does not account for — because an honest model is more useful than an overconfident one.

When forecasting is premature

If your historical data is unreliable, incomplete, or less than 12 months deep, a forecasting model will learn your data quality problems, not your business patterns. Clean your foundation first. We will tell you if that is the case — we would rather build something that works in month two than something that looks impressive in week one.

Diagnostic questions

When your forecast misses, do you know why — or does the team just adjust the number and move on? Does your planning process account for multiple scenarios, or does it depend on a single "most likely" projection? Could you explain to your board what assumptions your current forecast is built on?