No one shuts up about AI….we’re not either. As a planner, you need to know which products require your attention and why you are seeing spikes or declines in sales. That's the power of AI demand forecasting software. AI is achieving 24% increase in accuracy over traditional demand forecasting methods.

Don’t spread yourself thin in monthly planning meetings. Focus your attention on the products that matter most and make informed decisions that drive the business forward.

Our demand forecasting software uses advanced AI and machine learning algorithms to continuously update your forecasts and identify products that require attention. By leveraging the power of AI to adjust your forecasts throughout the year, you can achieve significant margin improvements that just aren't possible with traditional forecasting methods. Let’s dive in.

Where to focus?

In the current state, you set a baseline forecast in November for the following year. You have monthly planning review meetings to update the forecast. Considering you are already managing your forecasts by exception, is there a way to get help in this process?

At Granularity, we  see massive accuracy improvements by using AI to tell you where to focus your efforts in monthly planning meetings and in forecast adjustments. Which products are seeing the biggest changes from the baseline? Why are there differences? What’s happening in the market that you can’t see? By how much do you need to adjust forecasts? Ultimately, it is the ability to use AI to help adjust forecasts throughout the year that creates the biggest margin improvements.

The secret sauce: flagging products using AI

Flagged products are ones where the ML model has detected a market movement that was not accounted for in the baseline forecast. If we are able to adjust those products to the ML model, we can capitalize on changes in the market.

What’s powering the product flagging?  The product flagging process is powered by a continuous machine learning model that analyzes incoming data for each product and makes predictions about its market movement.

How does it work? The AI algorithm is trained to adapt to new data as it comes in, and is trained on a continuous flow of data, learning to more heavily weigh new indicators automatically. This means that the algorithm is constantly running forecasts for up to two months in advance. Our team is working towards training models on longer and longer forecast horizons.

Since the AI is continuously analyzing market trends, it's always ready to flag products that require attention and alert your planning team of potential risks.

Based on the sample products analyzed by the ML model, it achieves an accuracy 46 to 70 percentage points better in identifying market changes. That does not mean the baseline forecast is this far off for all products! The model is specifically highlighting the products that are seeing large swings in the market. By creating strategies to address these burning products, you'll see an increase in your overall accuracy and ability to move with the market.

Overall Benefits and Accuracy Improvements

Prediction for Feb (from Dec)

% improvement over baseline forecast

% improvement over ARIMA forecast

ML Model Forecast

Baseline  Forecast


Root Mean Square Error (RMSE)






Mean Absolute Error (MAE)






The ML model trained showed a significant improvement in forecast accuracy, with a 24% increase in RMSE and an 8% improvement in MAE over the baseline forecast. The data shows that by leveraging machine learning significant improvements in forecast accuracy are achieved over traditional methods like ARIMA (see assumptions at the bottom of the article). By continually analyzing incoming data and adjusting forecasts, the model stays ahead of market movements and raises risks to the demand planning team in a timely manner.

The estimated dollar benefit of the improved forecast accuracy depends heavily on the specific business and its inventory levels. Assuming a certain level of inventory, the improved forecast accuracy will help reduce stockouts or overstocking, leading to cost savings in inventory management. For example, if the improved forecast accuracy reduces stockouts by 10% and overstocking by 5%, and the average cost of a stockout is $250,000 per month and the average cost of overstocking is $150,000 per month, then the total cost savings would be $32,000 per month (10% x $250,000 x inventory level + 5% x $150,000 x inventory level). Additionally, the improved forecast helps optimize promotions, leading to cost savings in promotional spend. The specific dollar benefit would vary vastly depending on your inventory and production management practices.

Why is it performing better? The ML model performs better than the baseline forecast because it is able to capture complex relationships and patterns in the data that may not be apparent with traditional statistical methods. The ML model also has the ability to learn and adapt to new data, making it more flexible and robust in handling changes in the business environment. The ML model may incorporate more data sources, such as external factors like social media activity, holidays, and promotions, to improve the accuracy of the forecast.

Why isn’t ARIMA performing as well? ARIMA models are based on the assumption that the time series data is stationary, meaning that the statistical properties of the data remain constant over time. However, many real-world time series data, especially those from retail demand forecasting, are non-stationary, with trends and irregularities. We applied Auto-ARIMA to optimize the order of differencing, autoregressive, and moving average terms. However, we still found that some products are not suited for an ARIMA model, especially those with messy data over the past few years. More details on the ARIMA model will be described in a technical summary.

Ultimately, the best approach depends on the product. Some products work well for ARIMA, some require machine learning and deep learning approaches. However, by creating a mechanism for constantly updating forecasts using artificial intelligence and MLOps, we see an overall improvement in forecast accuracy.

This revolutionizes how businesses fulfill customer needs by proactively meeting customer demand, taking the customer experience to the next level.

How do you action the insights?

You may not be able to reduce orders you have placed. However, using early warning signs of demand means you can start activating our other levers.

  • Channel re-allocation — Reducing shipments out to stores, allocating more to eCommerce channels, etc. to proactively reduce operational costs
  • Pricing adjustments or promotional pricing — Competing on price early in the season rather than at end of season to move more inventory
  • Marketing positioning and campaigns — Diving deeper on what types of furniture will be excess and assessing whether a proactive marketing campaign or new positioning can help

The companies the are able to better move with the market and adjust their forecasts accordingly are the ones that will see the best profitability at year-end.

Take for example Bath & Body Works (not to be confused with Bed Bath & Beyond). Their process includes Weekly Planning Review meetings that include stakeholders from: the website engagement team, social media team, and ad spend team. They aggregate data on which products are being clicked and gaining popularity, which are performing lower than expected, and use social media discussions to understand what is going wrong. The demand planning team uses these direct weekly inputs to adjust strategies for promotional pricing and seasonal ordering, moving in tandem with the market.

Their ability to move with the market & their customers has consistently led to an operating margin that is 2 to 22 percentage points higher than their competitors. (Source: Stock Comaprison Source | MacroTrends)

Granularity is a web application where you can see predicted trends in the market for your categories. See aggregate trends from Tiktok, Instagram, Google, and more - and easily identify products flagged for growth or decline.


  • Baseline forecasting approach depends on the company. The baseline shown here represents a similar trend as last year.
  • The forecast is built on scaled product values, which are all scaled between 0 to 1. An RMSE of 0.139 is for scaled values between 0 and 1. Due to the scaling, the mean absolute percentage error (MAPE) is not one of the metrics used for comparison. The full methodology for outlier treatment and scaling will be shared in a technical summary.