While it is imperative for businesses to forecast company performance for future business decisions, traditional financial forecasting models can be limiting. These models take historical company trends and project them into the future, but they pull from a relatively small data set and can create potentially inaccurate assumptions. In contrast, artificial intelligence’s (AI) predictive analytics tools create adaptable, accurate forecasts that can take more variables into consideration than a traditional forecast.
Paro uses AI-based forecasting to model financial experts’ future earnings, the company’s projected sales revenue and overall future financial performance—with little error. To help your business optimize its forecasts, Paro’s AI/Data/BI Product Director, Ankit Tandon, breaks down the top do’s and don’ts for implementing AI forecasting in your business.
The benefits of AI forecasting over traditional financial forecasting models
When best practices are applied, AI-based forecasts can project future company financials with greater accuracy than traditional forecasts. This increased accuracy can allow company stakeholders to feel confident in making strategic decisions based on the predicted numbers. The key differentiators between AI forecasts and traditional forecasts are:
- The volume of data: Unlike traditional forecasts, AI and ML models aren’t limited in the amount, type or quality of data they can receive. As long as the data is available, AI forecasting models can consider both internal performance variables and external factors, such as macroeconomic conditions, stock market conditions and even how weather conditions impact business. Tandon states, “The limitation [of AI forecasting] isn’t what the model can work with, but what data you can give the model.”
- Adaptability: Whereas traditional financial forecasts must be manually recalibrated when circumstances change, AI-based forecasts are self-learning. Every time the algorithm is fed new data, the AI model will naturally adjust and reforecast company performance more accurately. The more real-time data the algorithm receives over a certain period, the higher the forecasting accuracy.
5 do’s of AI forecasting
When implemented correctly, AI forecasting models can achieve less than 5% error. Tandon stresses five key best practices when utilizing AI-based business forecasts.
1. Do go granular
Many companies use a top-down approach for forecasting, where they focus their model on top-line company financial projections. Paro champions a bottoms-up approach, where, as Ankit explains, “We start at the lowest possible level and then we sum that up.” Each company project gets its own AI forecast, and the sum of those forecasts equals the top-line performance. This approach not only increases forecast accuracy but also makes it easier to identify a culprit when performance doesn’t meet expectations.
2. Do retrain and monitor your AI
Tandon stresses that AI forecasts are only as good as the data you provide. If you don’t continuously provide the most up-to-date performance data, the model’s performance will degrade over time. Thus, it’s important to constantly retrain the model with the most recent data and then monitor performance in order to ensure continued accuracy.
3. Do keep a benchmark to detect anomalies
While AI models are good at adapting to normal performance fluctuations, they’re not as sophisticated at detecting anomalous events like a pandemic, earthquake, stock market crash or other disruptive event. An AI model could adjust your forecast incorrectly if it assumes that the anomalous performance occurred under normal circumstances.
As best practice, Tandon likes to keep a performance benchmark either in the model or on the side so that, if the forecast hits or exceeds said benchmark, the team will be alerted to the anomalous event and the data team can adjust the model accordingly.
4. Do maintain the human element
Internal financial analysts and business stakeholders know performance best. Consider using them to gut check an AI model’s results and determine whether the financial forecasts are reasonable. For example, a financial analyst who has been working at a company for years would likely know an AI model is off if it produced a 20% growth prediction, and the company has never had more than 2% growth. Tandon advises these internal experts to draw out the forecasting output themselves before using the AI forecast so that the team has a stick in the ground and can set the model in line with expectations.
5. Do define metrics before building the model
It’s just as important to know what the company is trying to measure with an AI model as it is to know how to build the model itself. It’s worth taking the time to align internally on what metrics you’re aiming to measure with your AI forecast and clearly define each metric. That way, it will be clear to both the data engineers building the model and the key stakeholders interpreting the model what the forecasted metrics are and how to use them for strategic decisions. “Before you can even build a model for it, you need to spend a lot of time debating the metric and all the edge cases for the metric, and how it’s calculated,” says Tandon.
5 don’ts of AI forecasting
As beneficial as AI and ML models can be for a company’s forecasting capabilities, they can also harm company performance if implemented or interpreted incorrectly. Here are some things to avoid when adding AI-based forecasts to your business processes.
1. Don’t use box products
Many companies will try to implement AI and ML forecasts by buying packaged AI products rather than hiring a data science team to build the model from scratch. While this might save the company time and onboarding costs, it may not work in the long run if the model doesn’t produce accurate results.
“AI is a huge buzzword in the industry, and a lot of companies will try to sell a product that’s like AI with the promise that it just works. But soon enough, there will be questions you have about the performance of it, and you won’t be able to answer them if you didn’t build it the right way from the ground up,” says Tandon.
Thus, Tandon recommends that companies build their AI-based forecasts internally in order to ensure they can explain and troubleshoot the nuances of its performance.
2. Don’t build more models than you need
Tandon stresses that AI-based forecasts require a significant investment in time, personnel, maintenance and data infrastructure. Companies ought to be wary of putting too much investment into an AI forecast that isn’t going to give them a positive ROI. In other words, a company shouldn’t invest more in developing an AI model than it could gain in incremental revenue for the next year once the model is implemented. Essentially, companies should do their due diligence to develop accurate and sophisticated AI models—but know when to stop.
3. Don’t expect 100% accuracy on day one
Tandon warns: “The hardest day in any AI implementation is day one, because when you launch, you will not get the performance that you are expecting, and if you do, you should be very suspicious of it.”
It’s extremely common for AI forecasts to only have 70-80% accuracy at first, but the model improves over time with fine-tuning and additional real-time data. Don’t give up on your model if you don’t see it work perfectly at first, and don’t set the expectation with company stakeholders that the model will work right away. Retrain, fine-tune and monitor for performance.
4. Don’t use AI to set targets right away
Every AI forecast has its quirks. For example, you may find post-launch that the AI algorithm is notoriously pessimistic about performance at the beginning of the month, so it initially forecasts monthly performance to be lower than in reality. These discrepancies can be adjusted and fixed over time. But until your team is confident that they’ve worked out all the quirks, don’t rely solely on the forecast to set your company’s financial targets. Instead, use the AI forecast as a starting place for your targets, and ultimately rely on human judgment to set the actual numbers your company will aim to hit.
5. Don’t build AI models in a silo
Building and maintaining AI forecasts doesn’t just involve one data scientist. There will be a team of engineers who build and adjust the model, a person or team that defines the metrics used in the model, analysts who interpret the forecasts, company executives who use the forecasts for strategic decision making and so on. In this sense, AI forecasting requires a multi-functional team with various responsibilities.
To avoid future confusion, Tandon reminds internal teams to define each member’s role in building, maintaining and interpreting the AI forecast at the outset of the project to account for all aspects of the AI. With a clearly defined cross-functional team, your AI forecasting project is likely to be a success.
Interested in elevating your company’s forecasts? Paro’s analysts and financial forecasting consultants can help you layer complexity into your forecasts, set granular business goals, build accurate assumptions and more. Request a free consultation to match with the best-fit fractional financial professional for your business.