Introduction
The pace of change in consumer preferences, seasonal fluctuations, and economic uncertainties can leave even the most experienced retailers struggling to keep up. Without the right tools, knowing when to stock up on holiday sweaters or when customers will shift to winter home goods can feel like guesswork or reactive analyses. Greater than retailers' need for robust consumer data is predictive forecasting models. This whitepaper outlines AnthologyAI’s retail forecasting model, illustrating how leveraging consumer insights collected ethically by our premier consumer app, Caden, helps retailers tackle these challenges head-on, enabling them to anticipate market shifts, optimize inventory, and strategically align their offerings in a constantly changing landscape.
We looked at sales trends across various categories including Apparel, Home products, Health, and Household Essentials. Our goal was to produce a reliable, data-driven model that reflects consumer behavior and aids in strategic forecasting for retail business decisions.
Data Preparation
For data selection, we focused on the high-impact categories from our detailed user profiles and their historical ordered items. (Apparel, Home, Health, etc.)
Feature Engineering
We focused on extracting pivotal user, spend, and retail dimensions from our extensive packet of over 500 features. Using dimensionality reduction techniques such as Principal Component Analysis (PCA), we distilled these down to the 100 most relevant features. This approach streamlined our dataset, focusing on information pertinent for model training while capturing monthly temporal trends from January 2022 to November 2024.
Model Selection & Tuning
We tried different models for time-series forecasting such as ARIMA, SARIMA, and Prophet. We back-tested the model on FRED (Federal Reserve Economic Data) to determine which model to select for forecasting. We chose ARIMA because the model had the lowest RMSE and the model prediction on the test set matched cyclical trends with the actual data.
After model selection, we normalized the data by dividing the average monthly volume by the distinct number of Caden users for each month. This ensures that all features are on a consistent scale, enhancing model performance. Before normalization, the model forecast exhibited a flat or declining trend. After normalizing the data, the model forecast reflected trends adjusting to user activity rather than absolute sales. The normalized forecast accounted for user fluctuations, ensuring comparability across months.
The Forecast
Apparel (Clothes, etc.)
Forecast Insights
Our model anticipates a sales peak for Apparel around February 2025, aligning with a consistent seasonal spike observed annually. Notably, a 22.50% increase is predicted for December 2024, followed by a steady moderation into early 2025.
- Seasonal Peaks: The normalized forecast shows a peak around February 2025, mirroring similar seasonal peaks observed over the past two years.
- Trends: The forecast captures a repeating cyclic trend, with peaks consistently occurring around the same months year over year.
Home (Bed, Pillows, etc.)
Forecast Insights
The Home category presents a predictable pattern of increased expenditure in January and February, a trend reinforced by past data. As projected, a post-peak decline follows, which our model accurately aligns with—highlighting its predictive robustness.
- Seasonal Trend: Normalized home spending consistently peaks during January to February each year.
- Post-Peak Decline: Spending declines in the months following this peak, forming a predictable seasonal pattern.
- Forecast Validation: The forecast aligns with this trend, projecting a similar January-February peak followed by a decline, consistent with historical data.
Health (Medicine, etc.)
Forecast Insights
Health-related spending historically surges during the winter months, peaking in March. Our forecasts adhere to this pattern, indicating a significant winter uptick consistent with findings from 2023 and 2024.
- Winter Spending Surge: Historical data from 2023 and 2024 indicate a consistent increase in health spending during the winter months, peaking around March.
- Forecast Alignment: The forecast reflects this pattern, projecting a steady rise in health purchases from December to March, maintaining alignment with past trends.
Household Essentials (Paper Towels, Toilet Paper, etc.)
Forecast Insights
A distinct seasonal spike in February mirrors patterns seen in preceding years. With a forecast showing a 61.09% increase in December 2024, the model captures these cyclic trends effectively.
- Seasonal Peaks: The normalized forecast shows a peak around February 2025, mirroring similar seasonal peaks observed over the past two years.
- Trends: The forecast captures a repeating cyclic trend, with peaks consistently occurring around the same months year over year.
Reflection
Based on our hypothesis, sales should be highest around the holidays (November/December).
The consistency of our findings with historical trends reinforces the value of our model. Evaluating its performance using RMSE values confirmed the effectiveness of the ARIMA model, validating our rigorous steps in model development, including dimensionality reduction through PCA and the normalization of data. These methods have significantly improved data quality and predictive accuracy.
Through AnthologyAI's proprietary data lens, we provide retailers with forecasts that transcend mere numbers; they offer actionable insights that empower businesses to adopt proactive strategies in dynamic consumer markets. By embracing data-driven decision-making, brands can enhance customer experiences, mitigate risks, and sharpen their competitive edge.
Ultimately, our forecasting approach transforms retail operations from reactive measures to strategic foresight, equipping businesses with the essential tools to navigate market complexities and drive sustained growth in an ever-evolving landscape.
