ARIMA Forecasting

The model is written ARIMA(p,d,q). The p term counts how many prior values enter the equation. The d term counts how many times the series is differenced, meaning each value replaced by its change from the prior period. The q term counts how many prior forecast errors get fed back in, which is what lets the model correct a persistent lean rather than repeating it.

Why differencing comes first

ARIMA assumes a stationary series, one whose average level and variance hold steady over time. Revenue at a growing company is the opposite of stationary. It climbs, and the size of its monthly swings climbs with it.

Differencing fixes the first problem. Modeling month-over-month change instead of the level removes the trend, and one differencing pass is enough for most bookings series. The growing-variance problem gets handled by modeling the log of the series, so a swing is treated as a percentage rather than a dollar amount. Skipping either step produces a model that fits the past and drifts badly forward.

SARIMA and the quarterly calendar

Plain ARIMA has no concept of a calendar, which makes it a poor fit for B2B SaaS on its own. Bookings arrive in a repeating shape. ORM sees Q2 and Q4 running stronger than Q1 and Q3 across its customer base, with the third month of each quarter closing more than the first two.

SARIMA handles this by adding seasonal terms at lag 12 for monthly data. The model then learns that a December spike is a December thing rather than evidence of acceleration. Without those terms, an ARIMA fit on SaaS bookings reads the quarter-end surge as momentum and carries it into the next month, which is precisely where quarter-start forecasts go wrong.

What ARIMA cannot see

ARIMA has one input: the history of the series. It cannot read pipeline created, win rate, or average deal size, so a change in the business reaches the model only after that change has already moved bookings. ORM identifies stale assumptions as the leading cause of forecast misses, and ARIMA is the purest example of a model built entirely on them. A competitor entering the market and compressing deal size will not appear in an ARIMA forecast until several months of compressed revenue have landed.

There is a second limit. ARIMA outputs an aggregate curve, not deal outcomes. It will tell you the shape of next quarter's revenue if conditions hold, but it cannot tell you which deals produce it, so it does not replace pipeline coverage analysis or a deal-level review. Treat a SARIMA projection as the statistical floor you compare the rep roll-up against, and track the difference across quarters as part of measuring forecast accuracy.