时间序列框架下含滞后Y变量的OLS模型LOOCV有效性咨询

Great questions—these are common pitfalls when mixing time series data with standard cross-validation techniques! Let’s break this down one by one.

Short answer: No, it's generally not valid.

Here's why: Leave-One-Out Cross-Validation (LOOCV) relies on the core assumption that your observations are independent and identically distributed (i.i.d.). But time series data is inherently sequential and dependent—each datapoint correlates with past (and often future) values. When you use LOOCV for time series, you remove a single observation at random (ignoring its position in the timeline) and train on the rest of the data.

This creates a critical look-ahead bias: your training set will include observations that occur after the test point. In real-world forecasting, you’d never have access to future data when predicting a past or current value, so this makes your error estimate overly optimistic. For example, if you’re testing the 10th monthly observation, your training set includes months 1-9 and 11-24. Using this model to predict month 10 gives you an error that understates your true out-of-sample performance, because the model learned patterns from future data that wouldn’t be available in practice.

Even more problematic—this approach is highly invalid, with amplified bias.

When your model includes lagged values of the dependent variable (e.g., using Y_{t-1} to predict Y_t), the look-ahead issue becomes even worse. Let’s say you’re testing observation t: your training set includes Y_{t+1}, Y_{t+2}, etc. While your model doesn’t directly use Y_{t+1} to predict Y_t, the model’s coefficients are estimated using this future data. This violates the fundamental rule of time series forecasting: you can’t use future information to predict the past.

Additionally, the serial correlation in time series means the omitted observation t is closely correlated with its neighboring datapoints (which are in the training set). LOOCV assumes the test observation is independent of the training set, which isn’t true here—this further distorts your error estimate, making it completely unreliable for assessing real-world performance.

Supporting References #

• Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction (2nd ed.). This textbook explicitly calls out that standard cross-validation (including LOOCV) fails for time series due to dependent observations. It recommends time-series-specific validation methods like forward chaining (walk-forward validation) that preserve temporal order.

• Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.). This practical forecasting resource dedicates a section to time series cross-validation, stating that LOOCV is inappropriate because it uses future data in training. It emphasizes that valid time series validation must never let the training set include data from after the test point.

• Box, G. E. P., Jenkins, G. M., Reinsel, G. C., & Ljung, G. M. (2015). Time Series Analysis: Forecasting and Control (5th ed.). A classic in the field, this book discusses the importance of respecting temporal dependence when evaluating model performance, noting that random cross-validation schemes (like LOOCV) do not account for the sequential structure of time series data.

内容的提问来源于stack exchange，提问作者Ozooha Ozooha