NeuralProphet

NeuralProphet is a time-series forecasting model that combines Prophet-style trend, seasonality, and event components with autoregressive and lagged-regressor terms fitted as a neural network in PyTorch.

NeuralProphet은 Prophet 방식의 추세·계절성·이벤트 요소에 자기회귀 및 시차 외생변수 항을 결합하여 PyTorch 기반 신경망으로 적합하는 시계열 예측 모형이다.

How do I forecast a time series with trend, seasonality, and autoregressive effects?

추세, 계절성, 자기회귀 효과가 있는 시계열은 어떻게 예측하나요?

NeuralProphet decomposes a series into trend, seasonal, holiday, and autoregressive components and fits them jointly with gradient descent, so each component can still be plotted and interpreted. It fits regularly spaced business or operational series, such as daily sales or demand, with clear seasonality and useful recent lags.

When to use NeuralProphet

• Daily, weekly, or hourly series with strong seasonal patterns
• Forecasts that should include holiday or event effects
• Short-term forecasts that benefit from recent autoregressive lags

When to use something else

• The series is short with simple structure: use exponential smoothing or ARIMA
• Many related series with complex nonlinear patterns: use global deep learning models such as TCN or N-BEATS

Assumptions and requirements

• Regularly spaced timestamps with enough history to cover several seasonal cycles
• Holdout or rolling-origin evaluation on the latest period
• Specification of the number of lags and forecast horizon
• Known future values for any future regressors

What the analysis reports

• Point forecasts with optional quantile intervals
• Component plots for trend, seasonality, and events
• Holdout error metrics such as MAE, RMSE, or MAPE
• Autoregressive weights for lagged values

How to report it

On the holdout period, the NeuralProphet model achieved an MAE of [value] and RMSE of [value] for a [horizon]-step forecast.

Related methods

• Prophet
• ARIMA Model
• Temporal Convolutional Network (TCN)

Run NeuralProphet on Quorant