DeepAR and Probabilistic RNN Forecasting
A recurrent neural network forecasting model that outputs a full probability distribution for each future time step, trained jointly across many related time series rather than one series at a time.
Prerequisites: Recurrent Neural Networks
Most classical time-series models are fit separately to each series, and many only produce a single point forecast rather than a sense of how uncertain that forecast is. DeepAR does both differently: it trains one recurrent neural network across an entire collection of related series at once (say, demand for thousands of related products), and instead of predicting a single number for each future time step, it predicts the parameters of a probability distribution (commonly Gaussian or negative binomial), giving both a forecast and a calibrated uncertainty band.
DeepAR trains a single recurrent network on many related time series simultaneously and outputs a full predictive distribution at each future step, rather than fitting one model per series and producing only a single-number forecast.
Why training across series helps
A single retail product's own history might be too short or sparse to fit a good model, but a network trained jointly across thousands of similar products can borrow statistical strength from all of them, learning shared seasonal and trend patterns while still conditioning its output on each specific series' own recent history and covariates (promotions, price, day of week).
It is widely used in demand forecasting and inventory planning, where knowing the range of plausible outcomes — not just the expected value — directly informs how much safety stock to hold.
Practice in interviews
Further reading
- Salinas, D. et al., 'DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks' (2020)