# Generate a Bottom-Up Forecast Bottom-up forecasting models analyze historical sales data at the individual SKU level to generate predictions. These models are excellent starting points for your business and intelligently learn from your sales patterns. As you continue to use Moselle, the forecasts will improve over time. While there are many factors to consider, the best indicator when choosing a model is your [MAPE score](https://learn.moselle.io/analytics/reporting/forecast-performance-report#how-to-read-ape-and-maps). Once you apply the model, you'll be able to evaluate its performance. *** ## Steps to Generate a Bottom-Up Forecast 1. On the [scenario plan flow](/planning-and-execution/forecasting/how-to-create-a-forecast.md), select the **Use AI Forecast Models** button. This reveals options for both Top-Down and Bottom-Up approaches. 2. Select the **Bottom-Up** approach and select one of the forecast models listed below. 3. After selecting your desired model, click **Next**. 4. Lastly, click **Generate Forecast** to incorporate machine learning projections into the planning page table. *** ## Available Bottom-Up Models | | Rome | Athens | Venice | Hong Kong | London | | --------------- | ----------------------- | ----------------------- | ------------------------------------------------------------------ | ------------------------------------------------------------------ | ----------------------------------------------------- | | Model Type | Statistical | Statistical | Probabilistic | Probabilistic | Deep Learning | | Cost / Compute | $ | $ | $$ | $$ | $$$ | | Forecast Method | Bottom-up | Bottom-up | Bottom-up | Bottom-up | Bottom-up | | Outcomes | Quick analysis, results | Quick analysis, results | Seasonal planning, layering in custom data i.e. Marketing Calendar | Seasonal planning, layering in custom data i.e. Marketing Calendar | Seasonality, retail stores (in addition to Ecommerce) | | Business Type | Simple business | Simple business | Fast-growing | Fast-growing | Complex, Enterprise | ### Rome Rome is built on [ARIMA (Auto Regressive Integrated Moving Average)](https://en.wikipedia.org/wiki/Autoregressive_integrated_moving_average), a widely-used statistical method for time series forecasting. This model blends multiple ARIMA forecasts, each designed to capture different time-series patterns like trends, seasonality, and short-term changes. **This model is optimized for:** * Substantial sales data, with regular sales activity * Predictable patterns that repeat over time (i.e. Seasonal Trends) **Example Use Case:** This model can predict moisturizer sales as they rise in winter due to dry skin and spike again in summer when SPF-infused products gain popularity. It helps identify high-demand weeks so teams can stock accordingly. ### Athens Athens is built on [Exponential Smoothing (ETS)](https://en.wikipedia.org/wiki/Exponential_smoothing), a statistical method that applies weighted averages to past observations, giving more importance to recent data. This model is designed for products with stable long-term trends and seasonality to create robust predictions. This approach can be effective for products with consistent historical patterns and minimal sudden changes. **This model is optimized for:** * Minimal no zero-sales weeks or months * Items with at least 12 months of data * Items with slow, steady growth or decline in sales * Predictable demand with few surprises **Example Use Case:** Your signature exfoliating cleanser has been slowly increasing over the past year, a notable steady growth trend, and with no surprises it stays steady week to week. ### Venice Venice is built on [Prophet](https://facebook.github.io/prophet/), an open-source forecasting tool developed by Meta. Prophet uses an additive model to fit non-linear trends with yearly, weekly, and daily seasonality, plus holiday effects. This model captures yearly, weekly, and holiday-based trends, and works especially well for retail products that see spikes in events such as Black Friday, back-to-school, or holiday sales. It helps you prepare for busy seasons by learning from past trends, even if some days of data are missing or sales were irregular. **This model is optimized for:** * Products with clear weekly and yearly patterns (e.g. Mondays, Black Friday) * Holiday or event-driven sales spikes * Sales with occasional gaps or missing data **Example Use Case:** Your essential t-shirt sales have a huge spike in November and December during the holiday season including a small bump in early September for back to school. Venice will track these patterns and recognize the holiday trends of those t-shirts from previous years. ### Hong Kong Hong Kong is Moselle's in-house built ensemble model. It blends three forecasting models into a single forecast using equal weighting. It evaluates patterns over time, seasonality, and messy data for a more balanced prediction. **This model is optimized for:** * Items with occasional zero-sales periods * Steady sales from week-to-week or month-to-month * Items with more than 2 years of sales history * Capturing seasonal patterns (e.g. Holidays, End of season sales, etc.) **Example Use Case:** A mid-tier skincare cream that sells steadily about 1,000 units per month with a noticeable uplift every December. Hong Kong averages its three component forecasts to provide a smooth projection that preserves the holiday bump while filtering out random week-to-week noise. ### London London is built on [Chronos](https://github.com/amazon-science/chronos-forecasting), an open-source family of pretrained time series forecasting models developed by Amazon. Chronos uses deep learning to transform time series data into token sequences, enabling probabilistic forecasts by learning patterns across diverse datasets. This lightweight deep learning model is designed for businesses needing accurate, short- to medium-term predictions with limited sales history. While it has a higher compute cost than statistical models, London delivers fast, reliable forecasts that adapt quickly to stable and seasonal demand patterns. **This model is optimized for:** * Items with short history (approximately 6 - 9 months) * Sales patterns that are simple and stable * Fast results needed for quick decisions * Online sales with a lot of historical data **Example Use Case:** You sell a jar of all-seasoned spices in your online shop every day which gives steady and predictable sales every day. Usually, you sell the same number each day with a small increase seasonally during summer, and sales do not jump around as much. *** ## How It Works Understanding the different types of forecasting models can help you choose the right one for your business: ### Ensemble An ensemble model combines multiple forecasting methods into one prediction. Think of it like getting opinions from several experts and averaging them together. By blending different approaches, ensemble models reduce the risk of any single method being wrong and typically produce more stable, reliable forecasts. Hong Kong is an example of an ensemble model that combines three different forecasting techniques. ### Statistical vs Probabilistic **Statistical models** (like Rome and Athens) use traditional mathematical formulas to identify patterns in your historical data. They work like a calculator that applies proven equations to your sales history. These models are fast, cost-effective, and work well when your sales patterns are consistent and predictable. However, they give you a single "best guess" forecast number. **Probabilistic models** (like Venice and Hong Kong) go a step further by providing a range of possible outcomes rather than just one number. Instead of saying "you'll sell 1,000 units," a probabilistic model might say "you'll likely sell between 900 and 1,100 units, with 1,000 being most probable." This range helps you plan for uncertainty and make better decisions about safety stock and inventory buffers. ### Probabilistic vs Deep Learning **Probabilistic models** use structured mathematical approaches that are designed to handle uncertainty. They're transparent in how they work and are great for capturing known patterns like seasonality and holidays. **Deep learning models** (like London) use artificial intelligence that learns patterns the way a human brain might—by processing vast amounts of data and finding hidden connections. These models can discover complex patterns that traditional methods might miss, especially when you have limited history or unusual sales behavior. The tradeoff is higher computational cost, but they can adapt quickly to new products or changing market conditions. *** ## Adjust Your Bottom-Up Forecast * Select the **Gear Icon** found in the top right corner, choose **Modify Forecast.** * Select **Change Forecast Method** * This will direct you to the **Forecast Creation Workflow** where you can re-select your **AI Forecast Models**. * Select **AI Forecast models**, and choose **Bottom Up**. This option will list you the available forecast models you can choose from. Once a model has been selected follow the Forecast Creation Workflow to finalize your Bottom Up forecast.
--- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://learn.moselle.io/planning-and-execution/forecasting/how-to-create-a-forecast/generate-a-bottom-up-forecast.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.