# How to explain global time series models to a customer?

For one of my customers I need to explain the concept of global models in simple words. Searching for simple introductions to the concept failed so far. All I can find are scientific studies, mostly about their advantage over local models.

What I need to explain is the basic functionality on how it works compared to local models. I know it is basically the same function, only that I am going to fit it to a group of time series instead of a single one. So if I have a set of 100 time series and want to predict the next 24 months, I need 100 iterations with local models (every column) and only 24 iterations (every forecast month) with global models.

How does this global forecast exactly happen, in terms of the data object and the loop function? I want to explain, why it is way faster than local models and why it is recommended to forecast on clusters of similar time series groups, to achieve higher accuracy.

Please share some insights or some good online references

If you want to reach outstanding results for your customer, you need both local and global models.

The customer should mainly understand the predictions' reliability, the confidence interval and the risks behind them.

There are algorithms that could estimate the predictions reliability. They could ease to select which model is the best in a specific situation.

Some local models could have better reliability due to seasonality or repetive behavior. Some global models could have better reliability due to correlations or anti-correlations with other variables (sometimes with a delay).

The prediction reliability algorithms give a score for each model, and the customer would use the one that has the better score.

Consequently, to maximize the benefits, you need to have local and global models firsts. Complex models that mix both things could be less effective and should be considered in a second phase.

Then, the interest of global models are quite obvious with concrete examples. Many real-world times series are related to weather or any other environmental influence.

It is always interesting to apply a correlation heatmap to show to the customers which are the correlated features and which are anti correlated, to select the most relevant ones. If you apply the correlation analysis to many features, you might detect surprising correlations that couldn't have been detected by a human. The ones that have no or little correlation could be discarded.

Some tools like dimensional reduction (t-SNE, UMAP, etc.) and explanatory white boxes might convince the customers about a global approach and could explain the correlations between variables.

In conclusion, an optimal approach would use several global and local models, using a reliability and explainability tools to increase the comprehension, take better decisions and improve models.

• Thanks a lot for your answer. I am actually currently working on a solution, that leverages the power of both, local and global models. In 90% of the cases, the global models did outperform the local approach so far. I used recursion on the global models, because they need to forecast two years of monthly data and recursion tends to give more constant estimates, which is desirable in case you do not know much about the distant future. However, I will still struggle to provide information, on how the forecast function works on panel data, compared to a single time series forecast object.
– LGe
Jan 27, 2022 at 14:53
• In that case, you should focus on explainability with tools like correlation analysis, white boxes or dimensional reduction. It would give more credibility and new opportunities, as you understand better the results. Risk and reliability algorithms are also very interesting to improve decisions. Many companies use story telling techniques (cf. storytellingwithdata.com) to help customers to understand complex models or results. Jan 28, 2022 at 9:35