If you're thinking about moving from spreadsheet forecasting to AI-driven forecasting, the first question is usually the same. What data do I actually need? The short version is six categories. The longer version, including the quality standards that separate a working model from a noisy one, is below.
Traditional forecasting leaned almost entirely on past sales. AI changes that. According to IBM research on AI demand forecasting, modern systems combine historical sales with supply chain data and external market indicators to sense demand before it even shows up on receipts. AWS Executive Insights notes that roughly 80% of today's supply chain data is now generated externally, which is why relying on internal sales history alone is no longer enough.
1. Historical sales data: POS and eCommerce transactions at SKU, store, and day level.
2. Product and assortment attributes: category, brand, size, colour, lifecycle stage.
3. Inventory and supply chain data: on-hand stock, lead times, supplier fill rates.
4. Pricing and promotional data: list price, elasticity, promotion calendars.
5. Customer and behavioural signals: search queries, page views, loyalty activity, sentiment.
6. External market factors: weather, events, macroeconomic indicators, competitor activity.
This is the foundation. Every forecasting model starts here, and it's usually the dataset most retailers already have in reasonable shape.
You need point-of-sale transactions at the SKU, store, and day level, along with eCommerce order data including baskets, returns, and cancellations. You also want channel-level splits so you can see how demand flows through in-store purchases, BOPIS, ship-from-store, and marketplaces separately. Most models need at least 24 to 36 months of history to properly capture seasonality and year-over-year patterns.
Key takeaway: Without clean, granular sales history, even the best model produces noisy outputs. If you're launching a brand new SKU with no sales history, product attributes take over the job.
Attributes are how AI forecasts demand for products that don't have a sales track record yet. Think new launches, seasonal items, or fashion SKUs that turn over every quarter.
The attributes that matter most are category and sub-category, brand, style, size, colour, pack size, and pricing tier. You also want to flag substitutes and complementary products, along with where each SKU sits in its lifecycle (introduction, growth, decline, or end-of-life). With these in place, an AI model can infer demand for a new lipstick shade by learning from how similar shades have performed across similar stores.
Key takeaway: Attribute-based modelling solves the cold-start problem, which is why retailers with frequent new launches see the biggest lift from AI forecasting.
Forecasts fall apart when they're disconnected from what's actually available to sell. You can't ship what you don't have, and you can't trust a forecast that ignores supply reality.
The essentials here are on-hand stock across stores, DCs, and in-transit inventory, along with supplier lead times and historical fill rates. You also want your available-to-promise (ATP) positions flowing into the model. Feeding this data in prevents the classic trap of forecasting demand you can't fulfil, and it's what enables demand sensing to catch present-day shifts instead of relying only on historical averages.
Key takeaway: A forecast without supply data is a wish list. Connect ERP, WMS, and POS before you turn the model on.
Price and promotion are the two biggest controllable drivers of demand, so leaving them out of a forecast is like trying to predict the weather without looking at pressure systems.
You want list price, discount depth, and price elasticity by SKU. You also want the promotion calendar with BOGO offers, markdowns, and flash sales, plus competitor pricing wherever available, and marketing spend broken down by channel and campaign. A good AI model isolates the lift caused by each promotion, so planners can predict demand under different pricing scenarios rather than guessing what a 20% off campaign might do.
Key takeaway: If your model can't explain promotional lift, it can't predict it. Historical promotion outcomes are the training data you need.
This is where modern forecasting gets interesting. A lot of the best predictive data sits upstream of the actual sale.
Website search queries, product page views, and cart additions act as leading indicators. A sudden spike in searches for a product can predict a demand surge days before POS data confirms it. Loyalty programme activity, customer segments, social media sentiment, trend signals, app engagement, and subscription renewal data all feed into this layer. When folded into a forecasting model, these signals catch shifts in customer intent while there's still time to react.
Key takeaway: Behavioural signals are leading indicators. POS data tells you what happened; search and browse data tell you what's about to happen.
This is what really separates an AI forecast from a spreadsheet forecast. External factors are everything happening outside your four walls that still influences what customers buy.
Weather forecasts at the store-location level, local events, holidays, school calendars, macroeconomic indicators like CPI and fuel prices, consumer confidence data, competitor activity, and category-level market trends all belong here. Gartner supply chain research consistently highlights external data integration as a top differentiator between leading and lagging planning organisations.
Key takeaway: Internal data tells you about your business; external data tells you about your market. Both are required.
These two terms get used interchangeably, but they're not the same thing. Here's how they compare across the dimensions that matter.
|
Dimension |
Demand Forecasting |
Demand Sensing |
|
Time horizon |
Weeks to months |
Zero to four weeks |
|
Primary data |
Historical sales, seasonality |
Real-time POS, web signals |
|
Refresh cadence |
Weekly or monthly |
Daily or real-time |
|
Best for |
Strategic planning, buying |
Replenishment, allocation |
Key takeaway: Most retailers need both. Demand forecasting sets the plan; demand sensing adjusts the plan when reality diverges from it.
Volume alone won't save you. Data quality is what actually drives accuracy, and there are a few baseline standards worth hitting before deploying anything.
Granularity: SKU, store, day for retail; SKU, customer, day for B2B.
Completeness: less than 5% missing values in core sales fields.
History: at least 24 months to capture full seasonal cycles.
Freshness: daily refresh minimum, ideally real-time for demand sensing.
Unification: a single source of truth across ERP, WMS, POS, and eCommerce.
Retailers who meet these standards typically see forecast error drop by 20 to 50% and lost sales from stockouts fall by up to 65%, according to industry benchmarks from McKinsey and IBM.
This is exactly the problem OnePint.ai is built to solve. OneTruth creates a single source of truth for inventory and sales data, while Pint Planning applies AI-powered demand forecasting and demand sensing on top of it.
Customers using the platform see 20 to 30% better forecast accuracy, 10 to 20% lower fulfillment costs, up to 85% fewer stockouts, and up to 15% higher sales. In practice, that has meant rebuilt inventory visibility for IPSY (the world's largest beauty subscription brand), inventory system modernisation for a major wholesale club, and predictive ATP and sourcing for a specialty jeweller. OnePint.ai was also recognised as a 2025 Gartner Cool Vendor in Supply Chain Planning Technology.
The platform integrates with existing POS, ERP, and eCommerce systems, so retailers can start forecasting with the data they already have, not the data they wish they had.
At least 24 months of SKU-level sales history, current inventory positions, and a price and promotion calendar. External data can be layered in progressively as your data maturity grows.
Yes. Attribute-based modelling uses product characteristics and the performance of similar SKUs to generate a cold-start forecast, even when there's zero sales history for the new item itself.
Daily for fast-moving retail categories, and real-time for demand-sensing use cases. Weekly or monthly refresh cycles are usually too slow for modern omnichannel operations.
Demand forecasting predicts future demand over weeks or months using historical and external data. Demand sensing uses real-time signals to detect shifts happening right now, usually within a zero to four week horizon. Most retailers need both.
Published benchmarks from IBM, AWS, and McKinsey consistently report 20 to 50% reductions in forecast error and up to 65% fewer lost sales from stockouts when AI forecasting replaces spreadsheet-based methods with clean, unified data.