Is "Average" Forecasting Killing Your Margins? Why you need adaptive forecasting

Most legacy planning tools—and yes, this includes the default demand planning modules inside major ERPs like NetSuite, SAP, or Microsoft Dynamics—are still stuck in the past. They cling to a single, static forecasting model, typically defaulting to a standard Moving Average or basic Exponential Smoothing (ETS).

But your catalog isn't a monolith. It’s a complex ecosystem of different behaviors. While this "one-size-fits-all" approach is acceptable for high-volume, stable commodities (like paper towels or basic white tees), it fails spectacularly when applied to the diverse reality of modern ecommerce inventory management. Today's retailers aren’t just selling staples; they are managing a chaotic mix of fast-moving trends, seasonal nightmares, new product launches with zero history, and noisy channels where sales spike for no apparent reason.

When you force a single mathematical logic onto these diverse behaviors, the result is damaging to your inventory turnover and cash flow:

→ Fast Movers get under-forecasted: When a product starts trending, a simple average moves too slowly to catch up. By the time the system tells you to buy more, you’re already out of stock.

→ Intermittent SKUs get over-forecasted: If you sell a spare part once a month, an average might tell you to keep steady stock, trapping cash in dead inventory that sits on the shelf for weeks.

→ Seasonal Items are mistimed: Because an "average" smooths out peaks and valleys, you end up replenishing holiday stock too late (missing the rush) or too early (clogging the warehouse).

The result? Teams inevitably resort to pulling data out of the system and into spreadsheets to manually override the bad math with human intuition. It’s risky, unscalable, and kills demand planning accuracy.

The Fix: The future of inventory planning isn't about finding the "perfect" algorithm; it's about using multiple algorithms and letting them compete.

That is why we are introducing Tightly’s Multi-Model Forecasting Engine. Instead of forcing a one-size-fits-all logic, our new engine treats every SKU as a unique case. Our new Adaptive Forecasting Engine is actually a multi-model, auto-selection engine that acts like a dedicated data scientist for every single SKU. It runs a background "tournament" between 7+ advanced models (including Prophet, Auto-ARIMA, and Croston) to find the specific mathematical fit for that product’s behaviour.

Here is a detailed look under the hood at how this machine learning supply chain technology works.

1. Runs Multiple Models

In the world of Time Series Analysis (the branch of data science dealing with time-based data), there is a fundamental rule: No single algorithm is good at everything. Every single sales history is made up of two things: Signal (the true trend) and Noise (random events, a rainy Tuesday, a bulk order). If you make a model very sensitive to catch a new trend (Signal), it will inevitably overreact to random spikes (Noise). If you make it stable to ignore the noise, it will be too slow to catch the trend. This is why both have to be taken into account.

Also, In machine learning for supply chain, we aren't just predicting numbers; we are predicting behaviors.

→Some behaviour is Habitual (Stable models).

→ Some behaviour is Emotional/Trending (Reactive models).

→ Some behaviour is Cyclical (Seasonal models).

Having this specific suite of models means Tightly isn't just "predicting demand: We have actually integrated a robust library of diverse statistical and machine learning temporal series algorithms. Depending on your specific plan, the engine can access "specialists" to handle your data:

• Moving Average: A Moving Average is perfect for stable products where time series data is flat and predictable.

• Auto-ARIMA & Auto-ETS: ARIMA predicts future values by looking at how recent data points relate to earlier ones. It uses past values to predict the next one (autoregression), removes long-term drift (differencing), and smooths short-term noise (moving average) to build the forecast. These advanced statistical methods work best for time series that show regular patterns, steady trends, and consistent timing.

• Prophet: We also use Prophet, the model developed by Meta. It is exceptionally good at handling "human" seasonality—like holidays and weekends—that trips up older AI demand forecasting tools. It builds smooth curves that follow predictable time-based behaviors.

• Croston/TSB: This is where we solve how to forecast intermittent demand. If you have products with many "zero sales" days, we use Croston/TSB. It predicts the probability of a sale rather than just an average quantity, which is critical if you want to accurately forecast slow moving inventory.

• Seasonal Naive: Perfect for products that repeat exact patterns year-over-year.

• Theta: A method that excels in spotting long-term trends.

The Solution is a Two-Phase Architecture:

To deliver demand planning accuracy at scale without crashing your system speed, Tightly therefore uses a smart two-phase workflow:

→ Phase 1 -  The system calculates a robust Moving Average baseline for all variants first. This ensures that every single SKU has a fallback forecast immediately.

→ Phase 2 - for products identified as complex or high-impact, the engine triggers the advanced model competition. Here, sophisticated models like Prophet and ARIMA compete against the baseline to see if they can offer a better prediction.

2. Intelligent Classification & The Art of the ABC

Before any math happens, the engine acts as a sophisticated filter. It doesn't just look at sales; it performs a dynamic form of ABC analysis inventory management. It analyzes the raw historical data of every product to understand its "personality," checking for data quality, seasonality patterns, and sparsity.

Based on this analysis, it classifies every product into a specific bucket:

• Baseline: Standard behavior suitable for simpler models.

• Intermittent: Products with gaps in sales history requiring specialized sparse-data models.

• Seasonal: Products with clear cyclical demand.

• Non-seasonal: Steady sellers that don't fluctuate with the calendar.

This automated classification isn't just for math; it’s a powerful tool for your SKU rationalization process. By identifying which products are truly intermittent versus steady, you can make smarter decisions about which items to keep and which to cut, streamlining your SKU rationalization efforts effortlessly.

3.  Smart Distribution Logic

How does Tightly decide which model wins? It relies on proof.

→ The 13-Week Validation Loop 

The engine performs a 13-week back-test validation. It effectively "hides" the last 13 weeks of actual sales data from the models and asks them to predict what happened. It then compares their predictions to reality. The model with the lowest error rate is crowned the winner. This ensures you are using the best demand planning software logic for that specific product, not just a generic guess.

Whilst Tightly stores the winning model, it also gives a transparent explanation for every product: a planner can inspect a SKU and see clearly: "This product is being forecasted using Croston/TSB because it was identified as intermittent." This transparency empowers operations teams to trust the data without needing to be data scientists.

→ 1-year forward weekly forecast

Once the winning model produces the forecast, Tightly generates a 1-year forward weekly forecast. However, a global number isn't enough—you need to know where to put the stock.

To solve this, the engine distributes that global forecast down to specific locations and sales channels using 2-year historical proportions. This ensures that even if your total demand is accurate, your inventory allocation across warehouses is equally precise.

Who Is This For?

This architecture is specifically designed for the machine learning supply chain of the future—tailored for mid-market to enterprise retailers with large catalogues (500+ SKUs).

.Complex Catalogues: Brands struggling to forecast slow-moving inventory alongside their best-sellers.

.Scalable Teams: Planners who need adaptive forecasting at scale but cannot afford the time to manually tune parameters for thousands of items.

.ERP Upgraders: Teams looking for a robust NetSuite demand planning alternative that handles real-world complexity without the enterprise price tag.

By moving to adaptive forecasting, you ensure fewer stockouts, less overstock, and a supply chain that reacts to the specific behaviour of every product you sell. It’s time to retire the "average" and give your business the precision it actually deserves. You build the brand. Let us handle the math. 

Schedule a demo to see the engine in action.

*starting Tier - Pro