Python中ARMA模型拟合与时间序列模拟的平稳性问题排查

Hey there! Let's walk through the issues in your code that are causing the degenerate simulation results and mismatched features compared to your original resid time series.

1. Critical Error in AR Coefficient Format for Simulation #

The biggest issue is how you're passing the AR coefficients to arma_generate_sample.

Statsmodels' ARMA model defines the process as:

$y_t = \phi_1 y_{t-1} + \phi_2 y_{t-2} + \varepsilon_t + \theta_1 \varepsilon_{t-1}$

But arma_generate_sample expects the characteristic equation coefficients for the AR part, which follow the form $1 - \phi_1 L - \phi_2 L^2 = 0$. That means you need to negate the AR parameters from your fitted model before appending the leading 1.

Your current code uses:

ar=np.append(1.,resid_model.arparams)

Which creates coefficients like [1, φ1, φ2]—this is backwards! Instead, you should use:

ar=np.append(1., -resid_model.arparams)

This gives the correct characteristic equation coefficients [1, -φ1, -φ2], which ensures the generated sequence follows the stationary AR process you intended.

2. Incorrect sigma Parameter Value #

Another small but impactful mistake: arma_generate_sample expects the standard deviation of the error term, but you're passing the variance (resid_model.sigma2). This will make the simulated sequence's variance much larger than your original resid series, leading to mismatched features.

Fix this by taking the square root:

sigma=np.sqrt(resid_model.sigma2)

3. Verifying Model Stationarity (Even with transparams=True) #

While transparams=True tells statsmodels to optimize parameters within the stationary/invertible domain during fitting, it's still good practice to verify the fitted model's stationarity to be sure. You can check the AR roots with:

print("AR roots:", resid_model.arroots)

For stationarity, all roots must have a magnitude greater than 1. If any root has a magnitude ≤ 1, your fitted model isn't stationary—this could happen if the underlying resid series doesn't actually follow a stationary ARMA(2,1) process, or if the fitting process hit a local minimum. In that case, you might need to test different ARMA orders or preprocess your resid series further.

Corrected Simulation Code #

Putting it all together, your simulation code should look like this:

import numpy as np
import statsmodels.tsa.api as smt

# Fit the ARMA model
resid_model = smt.ARMA(resid, (2, 1)).fit(
    maxlag=30, 
    ic='aic', 
    trend='nc',
    disp=False, 
    transparams=True
)

# Verify stationarity (optional but recommended)
print("AR roots (should all have |root| > 1):", resid_model.arroots)

# Generate correct simulated series
sim = smt.arma_generate_sample(
    ar=np.append(1., -resid_model.arparams),  # Negate AR params
    ma=np.append(1., resid_model.maparams),  # MA params are correct as-is
    nsample=tsteps, 
    sigma=np.sqrt(resid_model.sigma2)  # Use standard deviation
)

Why This Fixes the Degradation #

By correcting the AR coefficient sign, you're ensuring the generated sequence follows a stationary AR process instead of a divergent one (which causes the "degenerate trend" you saw). Fixing the sigma parameter aligns the noise variance with your original series, so the simulated data matches the resid features much closer.

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