能否用AR、MA、ARMA等统计模型分类时间序列？VAR模型输出异常

Hey there! I see exactly what's going on here—let's get your model to output those 0/1 labels you're expecting instead of floating-point numbers.

Why You're Getting Floating-Point Output #

First, let's clarify: the VAR model from the vars package is a vector autoregressive regression model, designed to predict continuous numerical time series. It estimates the conditional mean of your status variable (which happens to be coded 0/1 in your data) based on past values of all variables in the model. That's why it outputs continuous floating-point predictions instead of discrete 0/1 labels—it's treating status as a continuous variable, not a binary classification target.

Quick Fix: Convert Predictions to Binary Labels with a Threshold #

The simplest way to get your 0/1 labels is to apply a threshold to the continuous predictions. Since your status is 0 (no exercise) and 1 (exercising), a common starting point is a 0.5 threshold—adjust this based on your data's distribution or performance metrics like precision/recall.

Here's how to modify your existing code:

# After running your predict() call
prd <- predict(model, n.ahead = 10, ci = 0.95, dumvar = NULL)

# Extract the continuous forecast values for status
status_fcst <- prd$fcst$status[, "fcst"]

# Apply threshold to get binary labels (0/1)
# Adjust the threshold (0.5 here) based on your data's behavior
status_class <- ifelse(status_fcst >= 0.5, 1, 0)

# View your classification results
print(status_class)

Better Long-Term Solution: Use a Model Built for Discrete Time Series Classification #

If you want a more robust approach (since VAR isn't designed for binary targets), consider models tailored to discrete state prediction, especially since your problem is a state-switching scenario (exercise vs. no exercise):

1. Hidden Markov Models (HMMs)
   HMMs are perfect for modeling discrete states that switch over time, with continuous observations (your ax/ay/az data). Use the depmixS4 package to build one:

   library(depmixS4)
   
   # Build an HMM with 2 states (0 and 1) using your sensor variables as observations
   hmm_model <- depmix(list(ax ~ 1, ay ~ 1, az ~ 1), 
                       data = dt, 
                       nstates = 2, 
                       family = list(gaussian(), gaussian(), gaussian()))
   
   # Fit the model to your data
   hmm_fit <- fit(hmm_model)
   
   # Predict the hidden states (map state 1/2 to your 0/1 labels)
   hmm_predictions <- predict(hmm_fit)
   status_hmm_class <- ifelse(hmm_predictions$state == 1, 0, 1)
   

2. Discrete Vector Autoregressive Models
   For binary dependent variables, you can extend VAR with logit/probit links. Packages like mlogit or glmmTMB can help build these models, treating status as a binary outcome rather than continuous.

Final Notes #

• Test different thresholds if you stick with the VAR + threshold approach—use training data to find the threshold that maximizes accuracy, precision, or recall for your use case.
• HMMs will likely perform better for your state-switching scenario, as they explicitly model the discrete nature of your exercise status.

内容的提问来源于stack exchange，提问作者messi