Predictive modeling for finance

In partnership with

It was my 4th meeting with the predictive analytics team and I felt like we were going backwards.

“I see what the model says, but how am I supposed to explain it to the business partners?”

After 3 months and multiple iterations I turned from excited to frustrated.

The truth was that after years in the role, I was having trouble understanding the model myself - so of course the business partners were going to dismiss it.

My predictive analytics team was showing me what I affectionately refer to as a ‘shiny dashboard’ that was built on a complex statistical coding platform and ingested billions of data points every month. It supposedly was predicting future financial outcomes based on prior patterns of 150 unique data elements about each customer.

It looked AWESOME.

It was web-based, interactive, responsive, machine learning, and had all kinds of graphs.

All the buzz words and all the tech were included.

But it fell flat on one critical aspect:

The output was soooo close to answering a key question that I had, but it never quite landed.

The team had finished building the tool before I was brought into the project as the key stakeholder!

And that was the problem.

So what did we do?

I needed an answer to the business problem and I ran out of patience from my senior leadership and business partners.

So I had my best analyst build a simple predictive model in Excel that leveraged a few hundred rows of aggregated data, basic Excel formulas, and 3 data elements.

It was built to simply answer the business question by taking historical data and trending it into the future.

Guess what…

Everyone LOVED it!

We replicated that simple Excel model for the rest of the business units and financial results in a matter of weeks.

And we learned how to NOT deploy a predictive model.

In this week’s newsletter, we’re going to deep dive the approach you should take when building a predictive model to ensure it’s successful.

I’ve built dozens of them over my 11+ year career and they can be extremely rewarding when you get it right.

To be clear, a predictive model is not always crazy complex.

Sometimes they are simple trend analysis based on a few years of sales data by geography (for example).

As you read this post, think about a result that is hard for you to forecast accurately. Keep that in mind as your application for this week…

Learn AI in 5 minutes a day.

The Rundown is the world’s largest AI newsletter, read by over 600,000 professionals from companies like Apple, OpenAI, NASA, Tesla, and more.

Their expert research team spends all day learning what’s new in AI and gives you ‘the rundown’ of the most important developments in one free email every morning.

The result? Readers not only keep up with the insane pace of AI but also learn why it actually matters.

Sign up with one click.

You’re capping your finance career if you’ve been ignoring predictive modeling.

The more senior you get, the more you realize how critical the forecasting aspect of finance becomes.

Predicting the future and tying it to the P&L is one of the most valuable things we bring to an organization.

And the best way to predict the future is to build a model that does it for you.

But predictive models aren’t as scary as they sound if you have the right approach.

Here’s a high level guide on how to approach predictive modeling…
(next week we’ll go in depth into a few of my favorite methodologies with examples)

1. Start with a business problem, not a forecast methodology

As you get interested in predictive modeling, you’ll quickly feel like you are falling behind.

There are hundreds of tools and methodologies to learn! (ask ChatGPT for a list and it’ll look something like the list below):

• Time series

• Decision tree

• Random forest

• Machine learning

• Regression models

• Exponential smoothing

• And a bunch that I’ve never even heard of after doing this for 11 years!

I’d caution you to not fall into the trap of trying to learn a methodology and the fit it to a business problem.

You could spend a lifetime trying to master these methodologies.

Instead, you are better off identifying business problems and then working backwards to understand which methodology and tool is best to solve the problem.

This will ensure you don’t waste your time learning methodologies you don’t need.

And will keep you from overbuilding a predictive model with a fancy tool when all it needed was 1 Excel sheet.

2. How to identify a business problem

Therefore, the most critical part of predictive modeling is establishing your business problem.

For this exercise, I recommend you start with the parts of your P&L or forecast that you think ‘we should be able to predict this’.

The goal will be to narrow your predictive model down to solve a single variable/metric within your P&L or forecast.

Here’s a few common ideas:

• Customer churn

• Customer inquiries

• Sales trends

Using customer churn as an example, the business problems reads: “if we can better predict customer churn, then we’ll have earlier awareness of our sales needs to hit our revenue goal”

The typical problem with customer churn is that you can never know for sure if a recurring customer is going to buy again or stay with you.

But there are signs of someone that is about to churn (usually), especially if you know where to look.

It’s these signs and signals that finance can use to build a predictive model to feed your financial forecast for overall customer churn in the long term.

(next week’s newsletter will deep-dive a few different predictive modeling methodologies you can use to predict customer churn)

3. Look for external trends first

Often, the easiest and most impactful predictive models to implement come from tying external data to your internal company data.

For example: let’s say you identified that levels of theft within your stores increases in the summer when the temperature is warm (which happens to be true)

And since we can predict the average temperature with some degree of certainty (both monthly and on a daily basis), you will be able to build a predictive model to forecast theft trends based on weather trends.

Every business in the world is somehow connected to the macro economic trends of the industry they operate in.

Once a company hits a critical mass or large enough size of the market it operates in, then incorporating external trends into growth metrics becomes more applicable.

Think through how the labor market, cost of goods and services, or other miscellaneous expense trends are impacted by the external market.

Here’s a few random examples to get you started:

• Linking CPI of your industry to the cost of goods or services (Example: unit labor costs for full service restaurants)

• Incorporating underlying costs of goods sold into your gross margin projections (Example: medical supply and device manufacturing)

You’ll simply download the data from a site like FRED (both examples above), then trend that data along with your internal data to understand how they compare.

Then use your knowledge about the business to think through why they might not trend together perfectly - since they hardly ever do.

A word of caution for smaller companies…

For very new or small companies it is likely a waste of time trying to look for this connection especially as it relates to growth and revenue metrics.

Why?

Because a new company is more sensitive to gaining product/market fit and growing its share of the overall market.

Your time is better spent looking at the specific contracts and costs of goods that the business is involved with and getting intel from the business leaders on where they see those trends going over time.

4. Use internal data segmentation

Assuming you don’t have external trends to tie your internal data to for the business problem, you’ll need to leverage internal data to predict your key metrics.

This often relies on the segmentation of a specific metric within your P&L or forecast to draw out a better prediction.

For example: let’s say your software customers on the free tier submit 1.5x the number of support tickets on the weekends as your premium tier customers.

Your business problem you are trying to solve is: better predict customer support staffing needs in the short term to drive better customer experience and cost control.

If you have a sudden influx of free tier customers on your software, you can safely predict that your customer support needs will increase on the weekends.

This insight allows you to staff the team to fit the needs of the customer.

Knowing how your business partners think about the business is critical for this step.

If your business is a B2C software company then your business partners are likely not evaluating results based on geographic segments (e.g. states in the US).

Instead, the main segmentation that your business partners likely care most about are the types of customers (enterprise, non-enterprise) and where they are coming from (ads, organic traffic, outbound sales, etc.).

The key is thinking about each of those segments and asking ‘what data do I have that I think would show a unique difference between the customers for the business problem I am solving?’

Business problem + internal segmentation data = Predictive models that work

In Summary:

Predictive modeling is less about the methods, and much more about the business problem.

Look for external trends that mirror your internal ones. Then shift your attention to segmentation.

Next week we’ll dive deep into methods you can use to deploy your predictive analysis.

 Say hi 👋 on LinkedIn