Growing machine learning adoption in your finance org is tough. There are many levels to clear, with final bosses to beat. This happens most in aspects of finance that are already being done manually by humans, like forecasting (time series). Below are the three questions or “levels” you need to clear before any company can truly leverage machine learning to its full potential in finance for forecasting.
- Is the ML forecast as or more accurate than the current process?
- Can you explain the number generated by the black box?
- If ML forecasted $100, but actuals came in at $110, how do we explain the forecast variance?
Each level needs to be fully cleared before you can tackle the next. Let’s dive in.
First Level
The first level is pretty straight forward, you know what the accuracy bar is for the existing process, and in most cases you can create a machine learning forecast that can beat it. Even if you reach a similar level of accuracy with the ML process, ML can run in 95% less time than the manual process. Even then that can be a win for your finance team.
Second Level
Now you’ve built a machine learning process with great results. This is often where you will face the most resistance from your finance team. They will say “oh great the forecast is super accurate, but how do I know how it came up with its number?”. This is tougher than just creating an accurate forecast, since there is often a trade off between building the most accurate model and building the most interpretable one. Thankfully this is a hot area of research right now, with lots of great open source tools being released. The big key to interpretable forecasts for finance tasks is to understand the seasonality and trend of your forecast, and also how outside drivers (macro, internal KPIs) effect your forecast. Your CFO might not care that your model is mostly driven by historical growth rates over the last 2-3 years but telling her that a key driver in your model is the rise of interest rates and lowering consumer sentiment is a great way to tell a story around the forecast and get more people bought in.
Third Level
This is the final boss. This in uncharted territory because no data scientist or researcher is thinking about how their models impact the FP&A process at a company. This is a process specific to finance and one that doesn’t have a clear answer. If algorithms can be used to look into the future, then they could also be used to look into the past. Knowing how to automatically reconcile what was initially forecasted versus what actually happened would be a game changer. Allowing you to potentially automate much of the traditional close process in your finance team. Being able to explain the initial ML forecast in level two will allow you to eventually see the factors that contribute to the forecast variance during close.
Final Thoughts
I think about each one of these questions every single day. Open source tools that my team are working on, like Finn, are being actively improved to answer each one of these questions. If you’re interested in using these tools at your company or helping to make them even better, please reach out to me on LinkedIn.