Seasonal Product Sales Forecasting: How ImpactiveAI Does It

Every summer, sunscreen sales skyrocket, and in winter, moisturizing products fly off the shelves. Before Valentine's Day, gift set orders pour in, and when the new school year begins, demand for color cosmetics increases. Products that experience significant sales fluctuations during specific periods are called "seasonal products."

For retailers, forecasting seasonal product sales is an extremely challenging task. Prepare too much inventory, and you'll face losses from leftover stock. Prepare too little, and you'll miss sales opportunities. This becomes even more complex for health and beauty stores like Olive Young, which handle tens of thousands of products, each with different seasonal patterns.

ImpactiveAI uses cutting-edge AI technology to tackle this complex challenge of seasonal product sales forecasting. Rather than simply looking at historical data, we provide more accurate predictions by comprehensively analyzing various factors including weather, trends, promotions, and social media responses. In this article, we'll explore AI methodologies for seasonal product sales forecasting and apply them to predict seasonal product sales in a hypothetical cosmetics retail store.

‍

Advanced AI Methodologies for Seasonal Product Sales Forecasting

Time Series Data Collection, Processing, and Quality Assurance

Accurate AI predictions require high-quality, comprehensive data. This means integrating and preprocessing data from various sources including historical sales data, market trends, customer behavior, and external factors while maintaining strict data quality standards. To implement this effectively, robust data governance, automated data pipelines, and CDO investment are essential.

Simply put, AI needs quality data to make accurate predictions. Retail companies generate hundreds of thousands of transaction records daily, and finding meaningful patterns within this data requires systematic organization and validation processes.

‍

Time Series Decomposition (Trend, Seasonality, Residuals)

Time series decomposition is a core technique that breaks data into components - level, trend, seasonality, and noise - to identify patterns and improve modeling efficiency by isolating seasonality. There are additive and multiplicative models, with STL being particularly useful for handling evolving seasonality.

Time series decomposition is a technique that analyzes complex sales data by breaking it into several components. For example, looking at sales volume for products with strong seasonal characteristics, you can separate: ① overall upward trend, ② patterns that repeat every summer (seasonality), and ③ unpredictable random fluctuations (noise). This decomposition enables more accurate predictions for each component.

‍

Advanced Feature Engineering

Feature engineering is the process of transforming data so AI can better understand it. For example, just knowing it's "December" makes it difficult for AI to recognize it's Christmas season. However, providing additional information like "year-end season," "gift demand increase period," and "cold wave warnings" helps AI make more accurate predictions.

• Lag Features and Rolling Statistics: Lag features integrate past values of target variables or covariates into current predictions, providing historical context and capturing short-term dependencies and cyclical patterns. Rolling statistics (such as moving averages, maximums, minimums, standard deviations) are calculated across moving time windows to smooth noise, highlight local trends, and help capture temporal dynamics.
• Cyclical Encoding (Sine/Cosine Transformations): Time-based features like day of week, month, and quarter are essential for capturing cyclical and seasonal components. Sine and cosine transformations are particularly effective for representing inherently cyclical features (such as months of the year), ensuring models understand the continuity of cycles (December is followed by January).
• External Covariate Integration: AI models can integrate a wide range of external factors that significantly impact seasonal demand. This includes detailed weather data (temperature, precipitation, humidity), macroeconomic indicators (GDP, CPI, unemployment rate, oil prices), competitor activities (pricing, promotions, product launches), social media sentiment, and specific holiday or event indicators.
• Multi-modal Data Utilization for Rich Context Analysis: Cutting-edge AI makes comprehensive judgments like humans by combining various types of information. Rather than just looking at numerical sales data, it analyzes Instagram comments saying "This lipstick is amazing lately!", weather forecasts predicting "30°C tomorrow," and YouTube beauty influencer product review videos. By understanding and synthesizing different types of information, it can make much more accurate predictions than simply looking at past sales numbers.

‍

Advanced AI Model Comparison for Seasonality Forecasting

The AI models introduced in the table above each have different strengths. LSTMs excel at remembering long-term patterns, while Transformers comprehensively consider various external factors. Foundation models leverage pre-learned knowledge to predict new products well, and generative AI can simulate unexpected situations. ImpactiveAI combines these models according to specific situations to achieve optimal prediction performance.

‍

Probabilistic Forecasting for Quantifying Seasonal Demand Uncertainty

Unlike traditional single-point predictions, probabilistic forecasting explicitly quantifies the inherent uncertainty in demand by providing the entire probability distribution of possible future outcomes. This approach offers a more nuanced and reliable foundation for decision-making, revealing actual purchasing patterns and enabling risk-based inventory decisions across entire product portfolios.

This is particularly beneficial for seasonal products with highly variable demand curves, slow-moving items, and new product launches with limited historical data, helping companies manage risk more effectively. Instead of relying solely on single prediction values, companies can now optimize risk, desired service levels, and costs across various scenarios. This is especially important for seasonal products where the financial impact of excess inventory (waste, obsolescence costs) or stockouts (lost sales, customer dissatisfaction) can be substantial. Probabilistic forecasting restores confidence in the prediction process by explicitly acknowledging and quantifying uncertainty.

‍

Seasonal Product Sales Forecasting Example

We've analyzed methodologies for seasonal product sales forecasting so far. Health and beauty stores like Olive Young have many seasonal products whose demand fluctuates significantly based on various factors including seasonal changes, holidays, promotions, and beauty trends. Let's now briefly examine how ImpactiveAI predicts these sales.

‍

1. Goal and Scope Setting (What Will We Predict?)

• Define Target Products: Rather than predicting all Olive Young products at once, we first select specific product categories with distinct seasonality (such as summer suncare products, winter moisturizing products, holiday gift sets, specific seasonal limited edition products) or key seasonal products with high sales contribution.
• Forecasting Cycle and Period: Set prediction cycles such as weekly or monthly, and define short-term (next 4 weeks) or medium-term (next quarter) forecasting periods. This is adjusted to match business decision-making needs like inventory management and marketing campaign planning.
• Business Objectives: Clearly define specific goals to achieve through forecasting. For example, reducing stockout rates by 5%, cutting excess inventory costs by 10%, or increasing specific seasonal product sales by 15%.

For stores like Olive Young that handle diverse products, it's important to prioritize rather than forecasting all products simultaneously. We start with products that have high sales contribution or distinct seasonality, then gradually expand our scope.

‍

2. Data Collection and Advanced Feature Engineering (What Data Will We Use and How Will We Process It?)

High-quality diverse data is essential for accurate prediction. Data and feature engineering techniques considering Olive Young's characteristics include:

• Internal Data:
  • Historical Sales Data: SKU-level, store-level, daily sales volume and revenue data by date forms the most basic foundation for prediction. At least 2-3 years of historical data is needed, with long-term data including multiple seasonal cycles being particularly important.
  • Promotion and Discount Information: Large-scale discount events like "Olive Young Sale," buy-one-get-one-free events, specific brand discounts, including promotion start/end dates, discount rates, and marketing costs. This becomes a major cause of demand surges.
  • Inventory and Stockout Data: Historical stockout records are used as data to correct situations where sales volumes might be recorded lower than actual demand due to stockouts.
  • Product Attribute Data: Product categories (skincare, makeup, health supplements, etc.), brands, price ranges, new product launch dates, discontinuation dates, and other information.
  • Customer Data: Olive Young membership customer purchase history, preferred products, purchase cycles, etc., can be used for personalized predictions or micro-cohort forecasting.
• External Data:
  • Time-based Features: Extract basic time information like day of week, month, quarter, year, and week number. Particularly for cyclical data like months and weeks, apply cyclical encoding through sine and cosine transformations to help models learn continuous periodicity.
  • Holiday and Event Indicators: Create dummy variables (0/1) representing public holidays (New Year, Chuseok, Christmas, etc.), specific commemorative days (Valentine's Day, White Day, Children's Day, Pepero Day, etc.), and Olive Young's own regular/irregular promotional events (like Olive Young Sale).
  • Weather Data: Temperature, precipitation, humidity, etc., directly impact demand for specific beauty products like sunscreen, moisturizers, and sheet masks. Integrate regional weather data.
  • Macroeconomic Indicators: GDP, Consumer Price Index (CPI), unemployment rate, consumer sentiment index, etc., influence overall consumer psychology and purchasing power.
  • Social Media and Trend Data: Collect mention volume and sentiment analysis data about specific products, ingredients, and beauty routines from beauty-related communities, Instagram, TikTok, etc., to understand latest beauty trends and consumer interest levels. This is crucial for predicting demand changes in the rapidly evolving beauty market.
  • Competitor Activities: Information about major promotions, new product launches, price changes, etc., from competitors (such as Lalavla, Chicor, department store beauty corners) can influence Olive Young's demand.
  • Lag Features and Rolling Statistics: Add past sales volumes (such as previous week sales, previous month sales, same month previous year sales) as features, and calculate rolling statistics like moving averages and moving standard deviations to capture short-term patterns and volatility.
  • Multi-modal Data Integration: Integrate product images (such as packaging design trends), product description text, customer review text, etc., so AI models understand qualitative and contextual factors about products.

The data considered by Olive Young's AI prediction system is much more diverse than you might think. It analyzes not just historical sales data but also weather, trends, social media reactions, and competitor movements. This is because consumer purchasing decisions are influenced by very complex factors.

‍

3. Cutting-edge AI Model Selection and Training (Which AI Model Will We Use for Prediction?)

To predict Olive Young's complex seasonal demand, we consider the following advanced AI models:

• Deep Learning Models (e.g., LSTM/GRU): Effective at capturing complex seasonal patterns and long-term dependencies in long-term sales data (such as certain products consistently selling well during specific periods each year).
• Transformer-based Architectures (e.g., TFT): Leverage strengths in handling irregular data and seamlessly integrating various external covariates (weather, promotions, holidays, etc.). They can also provide probabilistic predictions to quantify prediction uncertainty.
• Time Series Foundation Models (e.g., Toto, Chronos): Utilize models pre-trained on large-scale diverse data to improve prediction performance, especially for new products or seasonal products with short sales history among Olive Young's vast product SKUs. This is because they have excellent generalization capabilities.
• Generative AI (GAN, Diffusion Models): Generate synthetic demand scenarios for situations not present in historical data, such as extreme promotional scenarios, unexpected beauty trend surges, or new seasonal product launches, performing "what-if" simulations. This helps prepare for unpredictable market changes and strengthen strategic planning.
• Causal AI: Goes beyond simple correlation to identify true causes of demand changes. For example, it separates whether the actual impact of a specific marketing campaign on sales was due to simple seasonal increases or the campaign's actual effect. This improves decision-making accuracy for marketing budget allocation and promotion timing decisions.
• Probabilistic Forecasting: Instead of single prediction values, provides entire probability distributions of possible demand to quantify prediction uncertainty. This is essential for making risk-based inventory decisions for Olive Young's diverse products (such as health supplements with expiration dates and color cosmetics with rapidly changing trends).

‍

4. AI Prediction System Implementation and Operations (How Will We Build and Manage the System?)

• Data Pipeline Construction: Build automated data pipelines that collect data in real-time or periodically from Olive Young's POS (Point of Sale) systems, ERP, CRM, and external data sources (weather service, social media APIs, etc.) and process it. Use Feature Stores to maintain feature consistency during training and inference, preventing "training-serving bias."
• Model Training and Validation:
  • Time Series Cross-validation: Instead of traditional cross-validation, use time series-specific cross-validation strategies like Rolling Window Validation to evaluate actual prediction performance of models.
  • Evaluation Metrics: Use metrics like MAPE, WAPE, sMAPE to evaluate prediction accuracy. For Olive Young specifically, since off-season product sales volumes can be close to zero, it's important to use sMAPE or WAPE, which are less sensitive to zero values. For probabilistic predictions, use specialized metrics like WQL and SQL.
• MLOps (Machine Learning Operations):
  • Continuous Monitoring: Continuously monitor the performance of deployed AI models to detect "Model Drift." Since beauty trends and market conditions change rapidly, we need to quickly identify when model accuracy deteriorates.
  • Automated Retraining: Build systems that automatically retrain and update models when new data accumulates or model drift is detected. This ensures prediction models always reflect the latest market conditions and consumer behavior.
  • Explainable AI (XAI): Apply XAI techniques to resolve the "black box" characteristics of complex deep learning models. This provides interpretable information about why prediction results turned out as they did (for example, "This month's sunscreen sales surge was due to temperature rise and specific influencer recommendations"), helping store managers and merchandisers trust predictions and use them for decision-making.

‍

Seasonal product sales forecasting has been a long-standing challenge in the retail industry. However, with ImpactiveAI's cutting-edge AI technology, we can now effectively solve this problem.

ImpactiveAI's differentiation lies in building AI that deeply understands business context beyond simple statistical prediction. We provide comprehensive forecasting that considers the impact of weather on cosmetics sales, mechanisms by which social media trends influence new product demand, sustainability of promotional effects, and more. Even in complex retail environments like Olive Young, we enable precise predictions at individual store and individual product levels, and quantify prediction uncertainty to support risk-based decision-making. This minimizes losses from stockouts or excess inventory while simultaneously improving customer satisfaction and profitability.

If seasonal product sales forecasting has been challenging for you, we invite you to experience innovative AI-based solutions with ImpactiveAI. We'll help you discover patterns hidden in data and accurately predict future demand, establishing a solid foundation for business success.

‍

‍