Picking a forecast algorithm without a data science team

Most FP&A forecasting in the wild is a linear trend with seasonal adjustments eyeballed by hand. That's fine for a stable business at 30% growth. It's not fine for a SaaS company growing 80% with a new product line and three pricing changes.

The right algorithm depends on your series. A few rules of thumb:

If the series is short (less than 18 months) and noisy, stick with linear regression or moving averages. The fancier models will overfit. Holt-Winters is a good middle ground — it captures trend + seasonality without needing a lot of data.

If you have 24+ months of clean monthly data with clear seasonality, ARIMA or SARIMA tends to win. They're slow to fit but accurate when the signal is there. The shortcut: run 5 algorithms side-by-side and look at the MAPE on the holdout set. The one with the lowest MAPE on hold-out (not training) is the one to lock.

If you're forecasting a new product line with no history, no algorithm will save you. Use a driver-based plan instead — model the inputs (price, units, retention) and let the math compose the revenue line. This is what NashOS's "Drivers + Member Formulas" surface is for.

Reading the metrics: R² close to 1 means the model fits the historical data well. RMSE / MAE in the same units as your series (dollars) tell you how much the typical forecast misses by. MAPE is a percentage — under 5% is excellent, 5–15% is good, 15–30% is workable, above 30% means the model isn't useful. Always look at MAPE on a holdout window, not on training data.

The last move: lock the winner. NashOS lets you pin the best-performing algorithm so the next forecast cycle uses it by default. You can revisit annually.