Treasurers and CFOs will feel increasing pressure to adopt AI technologies, such as machine learning and big data analytics, to facilitate financial forecasts. But with the available software tools, enriched data and, most importantly, the need for such initiatives, why isn’t this something that everyone is doing? It is irrational that humans refuse to relinquish more control to AI when running data analytics to produce a cash forecast.
When any structure is employed to design a forecast, there is always improvement over an unstructured funding amount based off sparse, random e-mails or bits of data. AI takes this structure to the next level. Humans have the ability to recognize just a few patterns. For instance, if a transaction called “payroll” appears every Friday for a certain amount, a human will pick up on it, remove the term “payroll,” and put in a random sequence of letters and numbers and it will be overlooked. This is where a machine will be able to find the repetition in different attributes of the payment. Ultimately, we are looking to have AI pick up where humans can no longer recognize the patterns.
The recognition of certain patterns does depend on humans to get things started, and this poses a risk to the model. When a human designs the model, there can be biases that affect the outcome, sometimes latent and sometimes for good reason. There may be an intercompany funding transaction that occurs quarterly, but for the sake of analysis, a treasurer may put in place parameters restricting the computer model to only the receiving side of the intercompany. When the funds are received, a flurry of payments will likely go out the door to meet top priority obligations, a pattern a computer will recognize. The corporation’s buildup of cash from operations will be unknown to this analysis, so the computer model will not be able to see how any of this cash going out the door benefits the company. This analysis may have other uses, but this human behavior, though a requirement, depicts how the human setting the parameters for AI forecast analysis may result in a flawed model.
Once your data is collected and you have your software, use machine learning to analyze the data. The technical formula for AI has only four main parts:
- Data clusters: This will be grouping transaction data by companies, or accounts.
- Interval determination: if you want the system to think in terms of days, weeks, months or quarters, or all of the above, you need to put these time intervals into the formula. This helps to find any seasonal trends or patterns that happen according to these time intervals.
- Weighing of intervals/clusters: This determines how important a set of data is for the overall analysis. Typically, this starts as an even weight (a retail company doesn’t experience all of its growth from Black Friday), but will evolve once insight is gained from where certain trends that are derived from a preliminary analysis predict the growth of a business.
- Limit calculations: As treasurers may know their business better than a machine, it is best to teach some of the “exceptional” events to the machine, so it does its job well. If you limit the focus of the analysis to the typical business activity and remove outliers such as a capital campaign funding receipt, the machine will be better at predicting typical business trends.
Now that you have taught the machine all that you know, the analysis will pull valuable insights. But this lesson does not stop here. Some very advanced programs will use the trends to find more trends and repeat this process to build an optimal business model. Most of the software available will present a resulting trend analysis and use predictive analytics to forecast a company’s activity. You can use the insights derived from this analysis to drill deeper into your data by redefining the four parameters above. If you notice a relatively flat trend, but you know one account in a company has large balance fluctuations, do a separate analysis on that individual account as the data cluster. Or if you do not see the spike in business from Black Friday, make sure you set your intervals to the day or week, not quarter or year.
FURTHER USE CASES
Aside from the core benefits already discussed regarding a quality forecast, we need to think of extraneous benefits when we have the power of AI in our office. In addition to identifying typical historical trends, using AI is excellent at identifying transactional patterns, which is useful for isolating transactions that do not align with those patterns. Many times, these atypical transactions are business as usual that can be explained by a quick call to a colleague in procurement, but other times they are fraudulent or mistakes that somehow made it through your approval workflow. When your forecasting is operating as it should be, these outlier transactions are easily identified.
In addition to pattern recognition, AI and machine learning are powerful tools for identifying relationships between data sets. For instance, if you’re a global retailer that needs cash on hand, you’ll want to choose an optimal amount to keep on premises. A decision to keep too much cash on-hand might not be an issue if it happens at a few stores, but if the mistake is being made globally, we could be talking about large sums. With machine learning, you can input data such as the geographic locations of your stores and the products being sold at each to see if the amount of cash on hand is warranted. These complex relationships would be near impossible to predict as an individual, but machine learning makes it possible.
Jason Dobbs, senior manager, and Kyle Olovson, CTP, senior consultant, are experts in corporate and international treasury at Actualize Consulting. For more information visit actualizeconsulting.com. They can be reached at firstname.lastname@example.org or email@example.com