Now is the time for CFOs and FP&A teams to transform how they forecast and get a little help from machines. Legacy planning systems have gone from behind the firewall to the cloud to more connected modeling. It is 2023. Where are the machine powered predictions to guide CFOs, FP&A teams, and budget owners to make better P&L decisions?
Machine learning (“ML”) is certainly not easy for back office functions. ML leverages massive amounts of financial and operational data to make useful predictions, integrating statistical reasoning to help businesses make repeatable, actionable insights. ML needs to be tested and iterated over and over and over again to improve the predictions. Additionally, purpose-built finance and accounting context is needed. And remember - ML is not one size fits all. Every enterprise’s business model is slightly different, which means the ML models need to be tweaked for each company.
A common misconception is that ML is driven by machines, when in fact, humans are behind the machines. When applying ML for P&L management, FP&A teams will be quarterbacking and holding the machines accountable, as a company’s business model evolves.
Below are tips for a successful ML deployment quarterbacked by FP&A teams based on what we have learned working with our Precanto customers.
Tip #1: Know what questions your enterprise wants answered
Before embarking on a ML adventure, identify what questions you want to have answered. How many FTEs do we need 3 years from now based on our revenue growth assumptions? How many and which of these “to be hired” FTEs can we realistically close before the end of quarter? Do we have enough recruiters and sourcers to achieve our hiring targets for this fiscal year? How do we optimize headcount expenses so our company can be non-GAAP profitable before the end of year?
Imagine you had a FP&A Magic 8 ball. What predictions would you ask your Magic 8 ball to make? The questions your enterprise needs answered are important to identify before deploying ML. Focus on where forecasting errors exist, where your largest spend categories are, or where most time is spent by FP&A teams to provide a forecast.
Tip #2: Identify what financial and operational data is needed
Once you know the predictions your enterprise wants, now you need to figure out what data you need. For a purpose-built ML model for FP&A, both financial and operational data from prior fiscal year periods at a granular level of detail will be needed.
Your enterprise resource planning (“ERP”) and enterprise performance management (“EPM”) solutions will have actual and planned financial data to leverage. Operational systems like your applicant tracking system (“ATS”) and human resources information system (“HRIS”) will house important data like salaries, start dates, bonus plans, and payroll information. If your organization is more sophisticated, an enterprise data warehouse (“EDW”) may exist to access both the financial and operational data necessary to set up your ML deployment.
Map the data needed back to the questions being asked and understand how the data will be accessed.
Tip #3: Understand what data needs to be cleaned, or what anomalies exist
Without clean data or a structured model, ML results in “garbage in, garbage out” similar to a poorly constructed financial model built in a spreadsheet by humans. Every enterprise has data silos. Every enterprise has bad data. Our experience is that ATS data, in particular, can be dirty. Similar to CRM data, ATS data is very transactional. Be patient as ATS data gets cleaned up because cleaner data means better (and more accurate) predictions. The data set won’t be perfect, but look for incremental improvements.
It is also important to discuss how enterprise anomalies should be addressed. For example, should the hiring freeze from 2 years ago be removed as an anomaly? How should you treat the COVID pandemic years in your ML model? How should the massive re-organization from last year be addressed in the ML model? You have options on how to address anomalies, what matters is that you discuss and test the anomalies as the enterprise ML models are built.
Make sure you get alignment on how to handle dirty data and anomalies before deploying your ML models. Remember your data will never be 100% cleaned with no anomalies.
Tip #4: Ensure the ML models are purpose-built with finance and accounting context
It is one thing to build a ML model, now try building a ML model that understands debits versus credits, operating expenses versus capital expenditures, run rates versus fixed costs, GAAP versus IFRS. Data scientists often come from computer engineering backgrounds, not having studied accounting, economics, or finance. FP&A and accounting teams need to bridge this gap with data scientists as ML models get built for P&L predictions. Come ready with questions and be prepared to challenge the ML models. Ensure the ML models are purpose-built with finance and accounting context that aligns with GAAP or IFRS accounting and that can be challenged by your CFO.
Tip #5: Validate the predictions of the ML models against prior forecasts
Select a previous forecast that was created with human built formulas and drivers that was loaded into your planning solution. Then, compare your human built forecast with the predictions from your purpose built ML models. How does the purpose built ML forecast compare with actuals for that period? What insights did the ML forecast derive to challenge any of the human built drivers? Instilling confidence in your FP&A team that the predictions work will allow you to move quickly to transform your legacy forecasting processes. The faster you kick the tires, the sooner you will get to P&L predictions to guide your enterprise to make better, more informed business decisions based on data.
Intrigued how you can quarterback change at your enterprise with P&L predictions? Come join the ML revolution and transform your legacy forecasting processes. Sign up to kick the tires of Precanto’s ML models.