Transcript
Transcript: The Grants and Contributions Model, by Housing, Infrastructure and Communities Canada
[00:00:00 Text on screen: AI-Powered Projects in the Government of Canada; The Grants and Contributions Model, by Housing, Infrastructure and Communities Canada ]
[00:00:03 Kirsten Gaudreau, Pierre Zwiller-Panicz and a colleague appear in video chat windows.]
Kirsten Gaudreau: Thank you Roxanne and hello everyone. Thank you for being here. Pierre and I are here...
[00:00:08 Kirsten Gaudreau appears in full screen.]
Kirsten Gaudreau: ...to present our machine learning forecasting model for grants and contributions. Next slide.
[00:00:15 Text on screen: Kirsten Gaudreau, Senior Financial Manager, Finance and Administration, HICC]
Kirsten Gaudreau: Let's begin. I'm going to be here to give you some context on the model, the problem we raised here at HICC.
[00:00:28 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, a cell phone, and the following text: A plan for AI Answers on Canada.ca, December 12, 2025.]
Kirsten Gaudreau: We will talk about the accuracy of the model, the impacts, the results, the implementation and the change management for the user.
[00:00:38 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: AGENDA. As described.]
Kirsten Gaudreau: Then, I'm going to hand the floor over to Pierre, who will give you a much more technical explanation of how the model works and how it was developed.
Next slide.
[00:00:48 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: The challenge: budget surpluses. As described.]
Kirsten Gaudreau: The project really started due to the challenge at HICC regarding budget surpluses in grants and contributions. For several years, we saw budget surpluses of over two billion dollars within our department. This led central agencies to request improved forecasts or budget requests for grants and contributions. What informs budget requests? These are our forecasts regarding grants and contributions. Previously, these forecasts were based on the cash flows received from the recipients.
But due to the truly unpredictable nature of infrastructure projects—we are talking about factors such as weather events or supply chain challenges. This means that improving cash flow as such is not always feasible.
So internally, we changed course and decided to focus on developing forecasting methods for each of our grant and contribution programs. So, these were really manual models, initially, that were based on historical data such as cash flows and actual expenses. But we quickly realized that we needed a standardized approach that could be applied to the majority of our current programs and that could also be adapted to future programs.
So, it was at that moment, within the HICC finance department, that we wanted to collaborate with the data office to build a ready-to-use machine learning model, a model that is scalable, based on historical data on grants and contributions. A model that is self-learning and adaptable to the current and future program. Next slide.
[00:02:40 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide with a chart titled: Insights on accuracy: machine learning performance.]
Kirsten Gaudreau: What we see on the screen here is the result of the model visualized in Power BI. This is what the end user sees, so, financial managers. What we see here are the results of the model. So, the green line, which is much closer to the actual spending in previous years, compared to the cash flow we receive from recipients, which is often overestimated. It is the managers who will use this tool to update our fiscal periods, and who will also use it to inform their conversations with clients. Next slide.
[00:03:23 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: Driving change: how machine learning makes a real difference. As described.]
Kirsten Gaudreau: Not only does the model lead to very good results, but it also allows us to better align our budget requests with our actual expenditures. We also saw a very positive impact in terms of internal workload. We were able to reduce our preparation time, the annual reference level updates, from three months to one month using the tool, and from one month to the next to update our financial situations. We were able to go from the three to five days that it previously took us to compile our manual methodologies to less than ten minutes using the model by clicking a button to update the data.
But the model has the potential to impact the federal government as a whole. So, if we think about the level of spending for the federal government, we have seen more than 150 billion in spending on grants and contributions for the government. Therefore, if other departments were to adopt the model, it could have a significant impact on the predictability of the department's financial framework. Even a small improvement in the forecasts could have a big impact on the financial framework. Next slide.
[00:04:38 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: Key considerations. As described.]
Kirsten Gaudreau: Now, for departments that may be considering adopting the model in their own context, there are certain considerations or key elements to consider. The first is that the model was built for the majority of programs. If, for example, in your context you have a program that is very different from the majority, the model might perform less well for that program. The second point is that the forecasts are based on data from existing projects. So, if a project is missing from your database or if some program data is incomplete, the model will not be able to make predictions for that missing information. The model also reflects the quality of the data. The better the data, the better the model's results.
Ultimately, the model is better suited to aggregate forecasts. What we mean by that is that the overall total performs better than the results at the project level. While information exists at the project level, these forecasts will balance out at the very end and lead to good results, either overall or perhaps even at the program level. With that, I will now hand the floor over to Pierre. Thank you.
Pierre Zwiller-Panicz: Thank you very much, Kirsten.
[00:06:11 Pierre Zwiller-Panicz appears in full screen.]
Now we're going to move on to the slightly more technical part of the presentation, since we hear a lot about artificial intelligence. But here, in this part of the presentation, we're really going to focus on what the model actually is and what's behind all these predictions.
[00:06:22 Text on screen: Pierre Zwiller-Panicz, Senior Data Science Analyst (Acting, HICC.]
Here on this slide, what you can see is a little bit of the pipeline that we use at HICC to be able to have the model from A to Z. The model is coded in Python. These results and outputs are put into our platform on Azure DevOps, which allows us to keep all the traces of the model's outputs. Next, we will visualize this in Power BI, especially for the finance teams. It's a very simple visualization tool, so financial officers don't necessarily need a lot of training. It's really very practical for them to use and especially to share; several financial analysts can be on the same Power BI at the same time. We found it to be much more practical for their day-to-day work. What you can see outlined in red, AI surveillance, is currently a framework that we are developing within HICC.
[00:06:25 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: Machine learning model pipeline.]
Because you need to know that when you put a model into production, it's not something you're simply going to launch and never configure or re-evaluate. On the contrary, we want to ensure that, once it is in production, we really keep an eye on it and make sure that the results do not get worse over time so that we can adjust them and maintain the accuracy that we currently have within HICC.
[00:07:44 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: Strategic selection of variables: A major asset for the use of machine learning.]
Pierre Zwiller-Panicz: Here, what you can see is basically all the questions we had when we started the project. We ended up with a database that is extremely comprehensive, with more than, I would say, 1,000 variables that we could take into account to build the model. It was definitely something we had to assess, especially at the beginning of the project, Kirsten and I. This random decision forest, which is our basic model, is efficient due to data that comes directly from our database, which you can see in blue. Here, in concrete terms, what we have is information at the program level, at the agreement level, and then at the project level.
It is really this layer of data that will allow us to have an overview and will allow the model to have an overview, primarily to give us more accurate results.
Based on this data, we created variables directly derived from it. The goal here is to inform the model of several things. Once the model takes the data into account, it will be informed of what has been spent on average, what has been spent in previous years and what remains to be spent in the project for future years. This will really give an indication and above all this will enable it not to reproduce the cash flows that we might receive from our recipients, which are usually overestimated. This will really allow the model to keep the amounts and the results in a relatively suitable bracket. But to truly understand what the model is, a random decision forest is basically a set of decision trees. Here, this is what you can see on the screen; it is the representation, let's say, of a decision tree. The way we like to illustrate this is to give, in terms of finances, is to say that a decision tree is equivalent to a financial analyst. I go see them at the beginning of the day and I tell them: All right, I would like to have a forecast on such and such a project. The analyst, with their information and knowledge, analyzes and tries to provide a result for this prediction.
[00:09:24 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: Mastering machine learning. As described.]
Here, for example, the analyst would look at what stage of the project we are at. Are we in the first year, the second year? What has been spent previously, or is left? Also, for example, the total duration of the project. All these parameters that I just mentioned are parameters that the random decision forest takes into account in its decision.
The idea here is to show a little bit of the decision-making process that can take place in a tree. But what causes problems, and the reason why we can not just have a single tree, is that if I give the tree a project, it will produce a result that will be relatively satisfactory. But the problem is that if I give it new projects that come in each year, then over time its results will start to decrease in accuracy. This is why we use a random decision forest, since a forest is a group of trees.
[00:10:47 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: Improving predictions: The transformative performance of algorithms of random decision forests. As described.]
Pierre Zwiller-Panicz: So now, when I come to give a project to my analyst, in reality, I am giving it to 200 analysts, who each have their own knowledge of the project, who maybe assess it differently and who can give me a result that is a form of overall result rather than a really very precise result. With this method and approach, we get results that are much more accurate over time, especially when we add new data to the equation, and knowing that in finance, we receive new data every day and that every year, we have agreements that change or projects that change. So, it is really important to have something that is truly reliable over time.
[00:11:32 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: Why are machine learning forecasts for G and C important? Plan proactively; Allocate budgets effectively; Prevent underspending or overspending; Improve transparency and control; Support strategic decision-making; Align funding with government priorities. As described.]
Pierre Zwiller-Panicz: These six points reiterate to a certain extent what we said above. We try to really predict proactively so we are not reactive to all these cash flows that we may receive. It helps us predict the budget and have an approach that is much more focused on data and our knowledge, rather than just knowledge. I think it also allows financial analysts to focus on large projects, which require much more work. This way we can run all the smaller projects, the more visible projects, programs that we know well, more easily with this model. This really allows them to focus their efforts on major projects.
[00:12:17 Split screen: Kirsten Gaudreau and Pierre Zwiller-Panicz in video chat windows, and a slide titled: Adopt and implement the HICC G and [illegible] forecasting model in your department. As described.]
Pierre Zwiller-Panicz: This is really the next step for our work. Since we have put the model into production and it is now working in HICC, our role now is to promote it in other departments. This was what we were doing with the Public Service Data Challenge that we were working on over the last few months. But here, this whole roadmap that we have built is really to show other departments how they could adopt the model and all the steps leading to adoption.
This is what our work is currently focused on. We will be presenting to the AGM committee in February. We are really keen to be able to expand this technology to other departments that might need it and for whom it would be effective. Thank you.
[00:13:02 Pierre Zwiller-Panicz appears in full screen.]
Pierre Zwiller-Panicz: We are reaching the end of our presentation. If you have any questions, feel free to ask them in the language of your choice.Thank you very much.
[00:13:12 CSPS logo appears on screen.]
[00:13:13 Text appears on screen: canada.ca/school.]
[00:13:17 Government of Canada logo appears on screen.]
