为何‘Self forecast’效果劣于‘Forecast from input’？LSTM时序预测疑问

Great question! This is a super common pain point with recursive time series forecasting using LSTMs—let’s break down exactly why this performance gap happens and what you can do about it:

First, Let’s Clarify the Two Approaches #

To make sure we’re aligned:

• Forecast from input: You feed actual historical time series values into the model at every step to predict the next value. The model always has ground-truth context to work with.
• Self forecast: You feed the model’s own previous predictions as input for the next step. It’s a closed-loop system where each new forecast depends entirely on the last one.

Key Reasons for the Performance Gap #

• Error accumulation: Every tiny prediction error in one step gets carried over to the next. Over time, these errors compound—like a game of telephone where a small distortion grows into something unrecognizable. Since the model never gets corrected with real data, the drift gets worse the longer you forecast.
• Distribution mismatch: Your LSTM was trained on real, observed time series data. When you use self-forecast, you’re feeding it synthetic values it never saw during training. The model isn’t optimized to handle this shifted data distribution, so its predictions degrade quickly.
• Weak long-term memory limits: Even well-tuned LSTMs struggle with very long sequences. Self-forecast requires the model to "remember" its own predictions over time, and if the model has too few units or insufficient layer depth, it loses track of the original pattern as steps pile up.
• Training-inference misalignment: If you used "teacher forcing" during training (feeding real data at every training step), the model never learned to recover from prediction errors. It’s used to having perfect input context, so when it has to work with its own flawed outputs, it can’t adapt effectively.

Quick Fixes to Boost Self Forecast Performance #

• Switch to direct multi-step forecasting: Instead of recursively predicting one step at a time, train the model to predict multiple future steps directly from the input sequence. This avoids error accumulation entirely.
• Try hybrid forecasting: Use real input data for the first few steps, then switch to self-forecast. This gives the model a solid ground-truth starting context before relying on its own predictions.
• Tweak your model architecture: Add more LSTM units, use bidirectional LSTMs (Bidirectional(LSTM(...))), or incorporate attention layers to help the model focus on critical historical patterns.
• Add error correction: If you have access to real values after some steps, feed the difference between the model’s prediction and the real value as an additional input feature. This helps the model adjust for past mistakes.

内容的提问来源于stack exchange，提问作者Roman