基于季节性时序数据预测每日投诉量：模型选型及是否限用回归咨询

Hey there! Let's break this down clearly for you since you're new to time series forecasting—no confusing jargon, just practical, actionable info.

1. Applicable Forecasting Models for Your Scenario #

Since you have seasonal daily data plus related sales records, here are the most suitable models, sorted by how easy they are for beginners to pick up:

Traditional Time Series Models (Great for Seasonality) #

• SARIMAX: This is the go-to classic for seasonal data. It builds on ARIMA (which handles trends and autocorrelation) by adding a seasonal component, and the "X" lets you include external features like your daily sales records. Perfect for your use case because it accounts for both the natural seasonality of complaints and how sales might drive them.
• Holt-Winters Exponential Smoothing: Part of the ETS (Exponential Smoothing State Space) family, this model is designed explicitly to handle trends and seasonality. It's simpler to set up than SARIMAX, making it ideal if you want a quick, interpretable baseline.

Machine Learning Models (Leverage Your Sales Data) #

• Prophet: Built by Facebook, this model was made for business-focused time series. It automatically detects seasonality (daily, weekly, yearly) and lets you plug in external features like sales. It's super beginner-friendly—you just feed it dates, complaint counts, and sales data, and it outputs forecasts with easy-to-understand visualizations.
• XGBoost/LightGBM: These are tree-based ensemble models that excel at capturing non-linear relationships (e.g., maybe a 10% sales spike leads to a 20% jump in complaints, not a linear 10%). You can add time-related features (like day of week, month, holidays) along with your sales data to make predictions. Most coding libraries have pre-built functions for these, so you don't have to start from scratch.

Advanced (For Later, Once You're Comfortable) #

• LSTM/GRU: These are deep learning models that can capture long-term patterns in time series. They work well if you have a lot of historical data, but they require more coding knowledge and parameter tuning. Save this for after you've mastered the simpler models.

2. Do You Have to Use Regression Models? #

Absolutely not! Regression models are just one small category of forecasting tools, and they're often not the best fit for time series with seasonality.

Here's why you don't need to stick to regression:

• The traditional time series models (SARIMAX, Holt-Winters) aren't regression models—they rely on the sequence's own patterns (autocorrelation, seasonality) instead of fitting a linear or non-linear equation to features.
• Tree-based models like XGBoost are non-linear and don't follow regression logic at all.
• Even if you did use regression, you'd have to manually engineer time features (like lagged complaint values, seasonal dummies) to make it work, which is more work than using a model built for time series.

Quick Tips for a Beginner #

• Start simple: Use Prophet or SARIMAX first to get a baseline forecast. They're easy to implement and will give you a sense of what's possible.
• Visualize your data first: Plot daily complaints over time to spot seasonal patterns (e.g., more complaints on weekends, holiday surges) and check how sales correlate with complaints. This will help you tweak your models.
• Use basic metrics like MAE (Mean Absolute Error) or RMSE (Root Mean Squared Error) to compare how well different models perform.

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