I$\text{'}$m working on time series forecasting and comparing different models: SARIMA, LSTM$,$ and ConvLSTM. I've noticed that for neural network models $($LSTM and ConvLSTM$),$ the results vary significantly with different random seeds. Here are my specific results:

ConvLSTM RMSE values with different seeds:

seed $42$: $67$

seed $41$: $220$

seed $40$: $83$

seed $39$: $147$

seed $38$: $104$

seed $37$: $70$

While my SARIMA model gives a consistent RMSE of $74.$

My questions are:

What is the proper way to compare the performance of traditional statistical models $($like SARIMA$)$ with neural network models that have varying results due to random initialization?

How should I determine which model performs better in this case?

What's the best practice for reporting results when comparing deterministic models $($SARIMA$)$ with stochastic models $($LSTM$/$ConvLSTM$)?$

For academic paper writing, which results should I use for comparison and how should I present them? Should I:

Use the best result from neural networks?

Use the average of multiple runs?

Report all results with statistical significance?

Or is there another recommended approach for academic publications?