How to Apply Gen AI & Machine Learning [2 x Frameworks]

Read time: 5 mins

If you’re unsure who to apply Generative AI & Machine learning as part of your finance activities, read on.

This is a special edition of the newsletter where I’m giving you 2 Frameworks.

But before we get into the detail, I’ll start by saying I’m not a data scientist.

Nor am I a developer.

But that doesn’t stop me putting data and code to use with a little help from AI.

Recently I had the privilege of presenting at GrowCFOs Global Finance Summit, and I got some pretty good feedback.

I used to play in metal bands, but I’d never been called a Rockstar after a presentation before! 🎸

What I wanted people to take away was this:

“Knowledge of what’s available now, so you can make best use of the AI tools available today, whilst allowing you to better prepare for the future.”

Machine Learning vs Generative AI

It’s easy to get confused about the different types of AI and their applications in finance.

So here’s a quick summary:

1. Machine Learning (ML) - ML allows you to deploy algorithms that will spot patterns in your data. You can use it to spot datapoints that don’t follow the trend (called outliers) or make predictions or forecast future trends using the patterns in historical data
2. Generative AI - Is a branch of AI that uses a sub-set of Machine Learning called ‘Deep Learning’ that uses MASSIVE amounts of data to train the model. It can then use this training data to create human like responses based on the themes it sees in it’s training data.

The uses case I focused on for Machine Learning was spend and revenue forecasting.

The use case I focused on for Generative AI was expense categorisation.

You’ll be able to find further resources at the end of the newsletter showing these examples.

Framework 1 - Machine Learning

1. Clean Data

Everything starts with data. I my example I took dummy data from Sage Intacct (supplier invoice data) and made sure it was formatted to include tidy dates, and monetary amounts. My revenue forecast was slightly different as this used random data that I generated with Python.

2. Load to DataFrame

Then I loaded my data into a DataFrame. I used AI to produce the Python code, and load it into the DataFrame using the Pandas Python library (for data formatting), ensuring the ‘Date’ column was recognised as a time series, and that the ‘Amount’ column was recognised as a Float (number).

3. Training the Model

I then got AI to tell me how to train the ARIMA model (the one it suggested for time series analysis). ARIMA requires a training step, before you can use it to create a forecast. But not all models require training. Prophet, which I used next, doesn’t require this training step.

4. Running the Forecast

You then need to run the code to actually create the forecast, but just running the forecast will not produce a visualisation.

5. Visualisation

I then used Matplotlb (for visualisation) and Plotly Dash (for dashboards) to visualise my data and create dashboards.

You’ll see this in the Presentation / Google Colab links I’ve provided at the end.

Framework 2 - Generative AI

The main problem with Generative AI is the variability in the responses you get. This means that asking the exact same question to an AI bot, will produce different answers every time.

1. Temperature

If you’ve used Microsoft Copilot, you might have seen the ‘Conversational Style’ which allows you to use precise, balanced, or creative response types.

These control what’s called ‘temperature’ in large language model terms. Higher temperatures are more creative, lower temperatures are more precise.

2. Provide Examples

If you can’t control your temperature, the best thing you can do is give examples. This is called either 0-shot (were you don’t give examples) or few-shot (where you give examples).

Here’s some expenses we might ask AI to categorise without examples.

Which might not produce very precise reults.

Whereas the below would produce much better prompts.

And now you can do this in Excel by using LABS.GENERATIVEAI in Excel.

3. Refine

Prompting is a bit of an art. So if you’re still seeing variability, continue to refine your prompts using either more, or clearer data examples to help train the AI in producing better outputs.

Putting it Into Practise

1. Check out the slides and commentary from the presentation here.
2. Listen to the audio of the presentation here.
3. Watch the video from the presentation here.
4. Check out the code and results I ran in Google Colab here.
5. Check out LABS.GENERATIVE AI here.

You’ll be an Gen AI & Machine learning pro in no time.

Until next time.

Adam

3 Ways I Can Help You

• Apply for the Finance Tech Accelerator - Transform your team's tech skills and harness the power of AI, Automation & Code in just 30 days, even if you're not techies. Apply here
• Become a Tech for Finance Community VIP - Access all areas pass to a community for finance pros looking to improve their performance with tech. Also includes weekly group coaching sessions to help with your challenges, discuss how to make the most of your existing systems, how to adopt new technologies, and much much more [3 Months FREE] Join here.
• Book a 1:1 coaching session - We'll turn your tech confusion into tech confidence. Start crushing it here