DeepAR: Probabilistic Forecasting with Neural Networks

So far, most forecasting models we’ve seen produce a single future value. DeepAR takes a different approach.

Instead of predicting one number, it predicts a probability distribution for every future time step.

The Real-World Problem

Imagine forecasting daily demand for a product across many stores.

• Some days sales are low
• Some days sales spike unexpectedly
• Uncertainty is unavoidable

A single prediction is not enough. Decision-makers need to understand risk.

What DeepAR Actually Predicts

DeepAR does not predict a single value like 120.

Instead, it predicts something like:

• Most likely value
• Range of possible values
• Probability of extreme outcomes

This is why DeepAR is called a probabilistic forecasting model.

Visualizing Sales Data

Below is a simulated daily sales series. It contains trend, weekly seasonality, and randomness.

How DeepAR Learns

DeepAR uses a recurrent neural network (RNN / LSTM / GRU).

At every time step, the model:

• Looks at previous values
• Learns temporal dependencies
• Outputs parameters of a probability distribution

Most commonly, it predicts the mean and variance of a Gaussian distribution.

Point Forecast vs Distribution

Below, the dark line shows the expected forecast. The shaded area shows uncertainty.

This visualization immediately answers:

• What is the likely future?
• How uncertain is that future?

Why Uncertainty Matters

In real businesses:

• Overstocking wastes money
• Understocking loses sales
• Planning needs confidence intervals

DeepAR gives decision-makers ranges instead of guesses.

Code Concept: DeepAR Logic

# DeepAR concept (simplified)

hidden = RNN(previous_values)

mu, sigma = Dense(hidden)

distribution = Normal(mu, sigma)

sample = distribution.sample()

Instead of predicting a value directly, the model predicts a distribution.

Confidence Intervals Explained

The shaded region in the plot represents confidence intervals.

• Narrow band → high confidence
• Wide band → high uncertainty

As we forecast further into the future, uncertainty naturally increases.

Where DeepAR Is Used

• Retail demand forecasting
• Energy consumption prediction
• Financial risk modeling
• Supply chain planning

Limitations of DeepAR

• Requires large datasets
• Training can be slow
• Harder to interpret than classical models

Practice Questions

Next lesson: Temporal Fusion Transformer (TFT) — combining attention and forecasting.