Abstract:Time-series forecasting is a critical challenge in various domains and has witnessed substantial progress in recent years. Many real-life scenarios, such as public health, economics, and social applications, involve feedback loops where predictions can influence the predicted outcome, subsequently altering the target variable's distribution. This phenomenon, known as performativity, introduces the potential for 'self-negating' or 'self-fulfilling' predictions. Despite extensive studies in classification problems across domains, performativity remains largely unexplored in the context of time-series forecasting from a machine-learning perspective.
In this paper, we formalize performative time-series forecasting (PeTS), addressing the challenge of accurate predictions when performativity-induced distribution shifts are possible. We propose a novel approach, Feature Performative-Shifting (FPS), which leverages the concept of delayed response to anticipate distribution shifts and subsequently predicts targets accordingly. We provide theoretical insights suggesting that FPS can potentially lead to reduced generalization error. We conduct comprehensive experiments using multiple time-series models on COVID-19 and traffic forecasting tasks. The results demonstrate that FPS consistently outperforms conventional time-series forecasting methods, highlighting its efficacy in handling performativity-induced challenges.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to make accurate predictions in the presence of distribution changes in time - series forecasting, since the prediction results may affect the future data distribution (the so - called "performativity" or "reflexivity" phenomenon). Specifically, the paper proposes a new method - Feature Performative - Shifting (FPS) - to deal with the distribution changes caused by prediction results, thereby improving the accuracy of time - series forecasting.
### Background of the Paper and Problem Definition
Time - series forecasting is an important research area and is widely used in many fields such as economics, urban computing, and epidemiology. Traditional forecasting models usually assume that the data distribution is fixed, but in reality, the prediction results may affect the future data distribution through feedback mechanisms. This phenomenon is called "performativity". For example, the prediction of traffic flow may affect people's travel decisions and then change the actual traffic flow; the prediction of diseases may affect the public's behavior and thus change the development trend of diseases.
Although "performativity" has been widely studied in classification problems, its application in time - series forecasting is relatively rare. Therefore, this paper aims to fill this gap and proposes a new framework to deal with the "performativity" problem in time - series forecasting.
### Problem Definition
The paper formally defines the **Performative Time - Series Forecasting (PeTS)** problem, that is, how to make accurate time - series forecasts in the presence of data distribution changes caused by prediction results. Specifically, the goal of PeTS is to optimize the prediction model \( \theta \) under the distribution change \( D(\theta) \) so as to minimize the prediction error:
\[ \theta^*_{\text{PeTS}} = \arg \min_{\theta} \mathbb{E}_{Z \sim D(\theta)} [L(Z; \theta)] \]
where \( D(\theta) \) represents the data distribution change caused by the prediction model \( \theta \), and \( L(Z; \theta) \) is the loss function, usually the mean square error (MSE).
### Proposed Method
To deal with the PeTS problem, the paper proposes the **Feature Performative - Shifting (FPS)** method. The main idea of FPS is to use the delayed response to predict future distribution changes, thereby improving the accuracy of prediction. The FPS method contains three main modules:
1. **Performative Time - Series Alignment**:
- Determine the appropriate delay time \( \tau \) so as to maximize the similarity between the delayed response and the target prediction.
- The formula is:
\[ \tau^* = \arg \max_{\tau} | \text{Similarity} (\{x_{t+\tau}\}, \{y_{t+k}\}) |, \quad \tau \in [0, k] \]
2. **Delay Translation Module**:
- Establish a mapping relationship from the current input features to the delayed response and train an autoregressive model using historical data.
- The loss function is:
\[ L_{\text{DT}}(f_\tau) = \text{MSE} (\{x_{\text{dr}}\}, f_\tau(\{x_t\})) \]
3. **Forecasting Module**:
- Use the delayed response and non - performative features to make the final target prediction.
- The loss function is:
\[ L_{\text{TS}}(g_\tau) = \text{MSE} (\{y_{t+k}\}, g_\tau(\