Temporal Fusion Transformer (TFT)

The Temporal Fusion Transformer is an attention-based deep learning model for multi-horizon time-series forecasting that combines recurrent layers, variable selection networks, and attention to use static, known-future, and past covariates.

템포럴 퓨전 트랜스포머는 순환층, 변수선택 네트워크, 어텐션을 결합하여 정적 변수, 미래에 알려진 변수, 과거 관측 변수를 함께 활용하는 다중시점 시계열 예측용 어텐션 기반 딥러닝 모형입니다.

How do I forecast several steps ahead using many covariates while still seeing which inputs matter?

여러 공변량을 사용해 여러 시점 앞을 예측하면서 어떤 입력이 중요한지도 알 수 있는 방법은 무엇인가요?

The Temporal Fusion Transformer produces multi-horizon quantile forecasts while using static attributes, past observed inputs, and covariates known in advance, such as holidays or planned promotions. It fits large forecasting datasets with rich covariates, and its variable selection weights and attention patterns give some insight into which inputs and time steps drive the forecasts.

When to use Temporal Fusion Transformer (TFT)

• Multi-step forecasts with covariates known in advance, such as calendars or prices
• Many related series with static attributes like store or product type
• Need for quantile forecasts with some interpretability of input importance

When to use something else

• Small datasets or a single short series: use exponential smoothing or ARIMA
• Many related series without rich covariates: use DeepAR

Assumptions and requirements

• Large training data across series or a long history
• Covariates classified as static, known-future, or past observed
• Chronological train-validation-test split
• Hyperparameter tuning and adequate computing resources

What the analysis reports

• Quantile forecasts for each horizon
• Variable selection importance for each input
• Attention weights over past time steps
• Holdout accuracy such as quantile loss or MAE

How to report it

The Temporal Fusion Transformer achieved a holdout [metric] of [value] at the [h]-step horizon, and [covariate] received the highest variable selection weight.

Related methods

• DeepAR Forecasting
• Recurrent Neural Network (LSTM, GRU, BiLSTM)
• N-BEATS

Run Temporal Fusion Transformer (TFT) on Quorant