- All three approaches can model complex seasonal sales, but they divide the problem differently.
- Prophet combines trend, Fourier seasonality, holidays, and optional regressors.
- DHR uses Fourier regression with ARIMA errors.
- MSTL + ETS first removes estimated seasonal components, forecasts the adjusted series, then adds projected seasons back.
- Use the table to compare the actual assumptions—not just the names.
| Concept | What it means | Northstar example |
|---|---|---|
| Representation is not the same as forecasting behavior. |
|
|
| Match strengths to the failure you need to fix. |
| If Northstar's adjusted residuals still have memory, DHR's ARIMA errors are a motivated candidate; that is a reason to test it, not declare it the winner. |
| Compare the full pipelines fairly. |
| Compare Prophet, DHR, and MSTL + ETS on the same Northstar forecast origins, horizons, calendars, and baselines, with every pipeline refit inside training windows. |
Prophet vs DHR vs MSTL + ETS
Compare Prophet, DHR, and MSTL + ETS by mechanics, assumptions, and use case—not invented performance scores.
The question it answers: Which trade-offs matter for Northstar's actual decision?
Compare the standard versions of these pipelines. Extensions and implementation defaults can change their behavior.
| Compare | Prophet | DHR | MSTL + ETS |
|---|---|---|---|
| What it does | Fits a trend/changepoint model plus Fourier seasonality, holiday effects, and optional regressors. | Fits regression with Fourier seasonal terms and optional known drivers; an ARIMA error process models remaining serial dependence. | Decomposes multiple seasonal patterns; fits nonseasonal ETS to the adjusted series; projects seasonal components and recombines. |
| Strengths | Convenient interface for trend changes, calendars, and holiday windows. A practical candidate for business series with strong patterns and enough history. | Explicit calendar/driver effects, smooth long or multiple seasonalities, and autocorrelated errors in one framework. Fourier orders control seasonal flexibility. | Readable component breakdown; can estimate evolving historical seasonal shapes. Modular: choose the adjusted-series model and seasonal projection separately. |
| Weaknesses | Trend extrapolation can be implausible; too much changepoint flexibility can overfit. Standard errors do not have an explicit ARIMA correlation model. Not automatically accurate. | Seasonal periods, Fourier orders, regressors, and error structure need care. Too many harmonics overfit; fixed seasonal coefficients may miss evolving shapes. ARIMA fitting can add cost. | Needs enough cycles and sensible smoothing settings. Boundary estimates can be fragile. Basic seasonal-naïve projection repeats the last seasonal shape; moving events need extra handling. |
| Trend & short-run memory | Trend is often piecewise linear, optionally with a saturating form. Standard Prophet does not explicitly model residual autocorrelation with AR/MA terms. | Trend may be represented by regression terms and/or the ARIMA error specification. Lagged values and innovations can capture remaining short-run dependence. | ETS forecasts level/trend behavior in the seasonally adjusted series. ETS innovations are not the same as choosing a general ARIMA error structure. |
| Changing seasonal patterns | Standard seasonal coefficients are fixed within a fit. Refitting or deliberately adding interactions/conditional effects can adapt the model. | Standard Fourier coefficients are fixed within a fit. Time-varying coefficients, interactions, or refitting are additional modeling choices. | Finite smoothing windows allow changes across the historical sample. The common future rule repeats final estimated cycles; it does not extrapolate new seasonal changes by itself. |
| Black Friday & moving events | Supply the actual holiday dates and lead/lag windows. The holiday interface is convenient, but the lift and future promotion assumptions still require evidence. | Add event-relative indicators or windows alongside seasonal waves. Calendar dates are known ahead; future discounts or spend must be known, forecast, or scenarized. | A fixed seasonal period is not an event calendar. Adjust/model the event separately using training data, forecast its effect, and add it back without double counting. |
| External predictors | Additional regressors can be supplied. Their future values must be available or supplied as scenarios. | A natural regression extension: price, promotions, weather, or availability indicators where justified. Unknown future inputs need forecasts or scenarios. | Plain MSTL + ETS has no general external-regressor layer. Incorporating drivers requires an adjustment or a different/hybrid forecasting pipeline. |
| Uncertainty | Intervals depend on model assumptions, fitting choices, and how future trend changes are simulated. Validate coverage; displayed bands are not guarantees. | ARIMA errors provide conditional forecast uncertainty. Fixed future inputs and fitted coefficients may omit predictor or parameter uncertainty; validate coverage. | Simple implementations often condition on estimated seasonal paths. Adjusted-series intervals alone can omit decomposition and future seasonal uncertainty; validate the whole pipeline. |
| What to tune and test | Changepoint flexibility, seasonal orders/regularization, and holiday effects. Check trend plausibility and event errors. | Fourier pairs for each period, driver specification, differencing and ARIMA orders. Check residual dependence and multi-step errors. | Periods, smoothing/robustness settings, ETS structure, and seasonal extrapolation. Refit decomposition at every backtest origin. |
Prophet and DHR both let you represent known calendar effects directly. MSTL + ETS remains a valid candidate if you deliberately add an event-adjustment layer. A sharp, moving sales event should not be left for smooth annual seasonality to discover automatically.
Same forecast origins, horizons, known-ahead inputs, preprocessing rules, and baselines. Check MAE/loss, bias, event-specific errors, interval coverage, runtime, and maintenance. There is no universal best model.
Fine print & assumptions
Decision guide: validate each candidate on the actual business decision and future-data tests.