ARIMA模型参数选择、警告及绘图报错问题排查求助

Let's break down each of the three problems you're facing and walk through practical fixes for them:

1. Invertible MA Coefficients Error (p=1, q=1) #

What's happening? #

The error ValueError: The computed initial MA coefficients are not invertible pops up because the estimated moving average (MA) coefficient for q=1 has an absolute value ≥ 1. ARIMA requires MA coefficients to be invertible (absolute value < 1) — this ensures the model can be reversed into a stable infinite AR series, which is critical for reliable forecasting.

Possible reasons include:

• ACF/PACF plots are a rough guide, not a strict rule — sometimes the "cutoff" pattern isn't as clear as it appears, leading to suboptimal p/q choices.
• Noise or small sample size in your data might be causing unstable coefficient estimates.
• The differencing order (d) might need adjustment, or a different p/q combination could work better.

Fixes to try: #

• Use automated parameter search: Skip manual ACF/PACF guesswork with pmdarima's auto_arima, which finds optimal (p,d,q) values while enforcing invertibility/stationarity. First install it with pip install pmdarima, then add this code:

  from pmdarima import auto_arima
  stepwise_fit = auto_arima(data_log_diff, trace=True, suppress_warnings=True)
  print(stepwise_fit.summary())
  # Build your ARIMA model using the suggested order
  

• Test nearby q values: Try q=0 (which you know works) or q=2 to see if the model runs without errors.
• Recheck stationarity: Even if your Dickey-Fuller test passed, ensure there are no hidden trends/seasonality that could destabilize estimates.

2. No Frequency Information Warning #

What's happening? #

The warning ValueWarning: No frequency information was provided, so inferred frequency MS will be used means your pandas DatetimeIndex lacks an explicitly defined frequency. Statsmodels relies on this metadata for time-based operations (like differencing and forecasting), so it's guessing your data is monthly (MS = Month Start).

Quick fix: #

After setting your index, explicitly define the frequency to match your data (since the warning infers MS, we'll use that):

indexed_data = data.set_index(['Month'])
# Add this line to set frequency explicitly
indexed_data = indexed_data.asfreq('MS')

Alternatively, specify the frequency when converting the Month column to datetime:

data['Month'] = pd.to_datetime(data['Month'], infer_datetime_format=True)
indexed_data = data.set_index(['Month']).asfreq('MS')

This eliminates the warning and makes your time series index more robust.

3. Timestamp vs Integer Addition/Subtraction Error #

What's happening? #

The TypeError occurs because of an index mismatch between data_log_diff and results_ARIMA.fittedvalues. When you fit an ARIMA(1,1,0) model, the fitted values start from the second observation (due to differencing requiring a lag), but older statsmodels versions might not align these indexes correctly. This leads to a conflict when plotting, as the code tries to mix Timestamp and integer operations.

Fix by aligning indexes: #

Modify your plotting code to ensure both series use matching indexes. You can either slice the fitted values or use the predict() method with explicit bounds:

# Option 1: Slice fitted values to match data_log_diff's length
plt.plot(data_log_diff)
plt.plot(results_ARIMA.fittedvalues.iloc[1:], color='red')  # Skip the first mismatched value
plt.title('RSS: %.4f'% sum((results_ARIMA.fittedvalues.iloc[1:]-data_log_diff)**2))

# Option 2: Use predict() to generate perfectly aligned values
preds = results_ARIMA.predict(start=1, end=len(data_log_diff))
plt.plot(data_log_diff)
plt.plot(preds, color='red')
plt.title('RSS: %.4f'% sum((preds-data_log_diff)**2))

This ensures both series use the same Timestamp index, resolving the integer-Timestamp conflict.

内容的提问来源于stack exchange，提问作者Pavan