We'll get back to you within 24 hours.
We built an AI demand forecasting and replenishment system for a 40-store grocery and FMCG retail chain — predicting demand per SKU per store per day with 94% accuracy, auto-generating purchase orders, and freeing ₹3.2 Cr in trapped working capital.
Industry Context
India's organised retail sector is undergoing a structural transformation. FICCI-Technopak research projects the organised grocery and FMCG retail market to grow from ₹4.4 lakh crore in 2023 to ₹7.8 lakh crore by 2028 — a 12% compound annual growth rate driven by rising incomes, urbanisation, and the rapid expansion of modern trade formats into Tier 2 and 3 cities. Retail chains that manage inventory efficiently during this expansion phase will capture market share permanently; those that don't will see their working capital trapped in excess stock while simultaneously losing customers to competitors who can reliably stock what buyers need. The perishable waste problem in Indian grocery retail is severe: the industry average for perishable waste runs at 15–22% of perishable inventory value monthly, a figure that dwarfs comparable benchmarks in mature retail markets (3–5% in UK grocery). This gap exists primarily because Indian retail has historically operated on intuition-based ordering — store managers buying based on last week's sales, seasonal memory, and supplier relationships rather than data-driven forecasting. The stockout problem is equally damaging: consumer behaviour research shows that when a customer encounters a stockout on a preferred brand, they switch to a competitor brand 60% of the time. For premium FMCG categories, stockout brand-switching is even higher at 68–72%. Every stockout event is not just a lost sale — it is a potential permanent customer transfer to a competing brand. AI demand forecasting closes both gaps simultaneously.
Indian organised grocery retailers waste 15–22% of perishable inventory value monthly due to over-ordering driven by gut-feel buying decisions. This waste rate is 4–5x higher than mature retail markets because Indian retail lacks the data infrastructure and analytical capability to forecast perishable demand at granular store and SKU level. AI forecasting directly targets this waste — the 6.4% waste reduction achieved in this case study represents the largest single cost saving the client had ever realised from a technology investment.
The Nielsen research finding that 60% of shoppers switch brands permanently after a stockout is one of the most commercially significant statistics in retail. For a 40-store chain carrying 4,500 SKUs, 230 daily stockout events represent a continuous, compounding brand loyalty erosion. AI forecasting reduces stockouts by maintaining optimal stock levels — not by over-ordering (which creates waste) but by ordering the right quantity at the right time based on granular demand signals.
Over-ordering creates a second financial problem beyond waste: working capital trapped in excess inventory earns no return and reduces the retailer's ability to invest in growth. For a 40-store chain with ₹26 Cr in inventory across stores, even a 12% reduction in average stock holding (through more precise ordering) frees ₹3.2 Cr in cash. At a 12% cost of capital, this working capital release generates ₹38.4 lakhs in annual value even before counting the waste reduction savings.
The Challenge
Our Solution
A store-level, SKU-level daily demand forecasting system that accounts for seasonality, promotions, weather, and local events — automatically generating purchase orders and flagging expiry risks.
The forecasting engine produces 180,000 daily demand predictions — one for each of the 4,500 SKUs across all 40 stores. This granularity is what makes the system genuinely useful: a chain-level forecast that says "we'll sell X units of Brand Y biscuits this week" is far less actionable than a store-level forecast that says "Store 12 in Andheri will sell 47 units of Brand Y biscuits on Thursday because it's near an office complex and people buy snacks on Thursdays, while Store 31 in a residential area will sell only 12 units." The LightGBM model handles the tabular demand prediction for non-perishable and slow-moving SKUs, capturing complex feature interactions (day of week × weather × proximity to payday × promotional status) that simpler statistical models miss. The Facebook Prophet component handles the time series decomposition for products with strong seasonal patterns — fresh produce, dairy, festival-linked categories — where trend, weekly seasonality, and annual seasonality need to be modelled as explicit components rather than learned from feature interactions.
Demand prediction accuracy is fundamentally limited by the quality of the input signals, and the most common failure mode in retail forecasting is using only historical sales data without accounting for the external factors that cause demand to deviate from historical patterns. Our system integrates four categories of external signals: public holiday and regional festival calendars (Diwali, Navratri, Eid, Christmas, and 340+ regional holidays and fasting days that affect specific product categories); hyperlocal weather forecasts via the OpenWeatherMap API (ice cream demand rises 23% for every 5°C above 32°C average daily temperature; umbrella demand spikes 48 hours before a forecast heavy rain event); the retailer's own promotional calendar (price promotions, BOGO offers, display placements) integrated from the central ERP system; and neighbourhood event data (cricket matches at nearby grounds, local festivals, school exam periods that reduce evening footfall). Each external signal has a learned demand elasticity coefficient, trained from historical instances where the signal was present and demand deviated from baseline.
The most operationally transformative component of the system is not the forecasting model itself but the automated purchase order generation engine that converts forecasts into actionable replenishment decisions. Every morning at 6 AM, the system computes for every SKU at every store: current stock level (from POS real-time stock tracking), forecast demand for the next 14 days, supplier lead time (stored per SKU from historical delivery records), minimum order quantities (supplier constraint), and economic order quantity optimisation to minimise ordering frequency. From these inputs, it generates draft purchase orders for the next 3 days — a rolling 3-day horizon that balances advance planning with forecast accuracy (24-hour forecasts are 94% accurate; 14-day forecasts are 82% accurate). Buyers receive the draft orders in a prioritisation dashboard sorted by risk: orders where stock will reach zero before the next delivery arrive first, followed by expiry risk items, then routine replenishments. Routine orders within ±20% of the AI suggestion are auto-approved without buyer review; exceptions and anomalies are flagged for human judgment. This process saves each buyer 3 hours daily that previously went to manual PO construction.
For perishable categories — dairy, fresh produce, bakery, chilled meats — the system runs a parallel expiry risk model that projects, for every batch of perishable inventory in each store, the probability of reaching expiry date before selling through. When an item crosses a configurable risk threshold (default: 60% probability of expiry within 5 days), the system triggers one of three automated interventions depending on how much time remains: a markdown promotion alert pushed to the store manager's app for items with 5+ days to expiry (discounting typically moves perishable stock within 24–48 hours when well-executed); an inter-store transfer recommendation for items where a nearby store is undersupplied and can receive the stock before expiry; or an urgent discount alert for items with 2–3 days to expiry where immediate price action is required. In the first 6 months of operation, the expiry dashboard-triggered interventions saved an additional 4% of perishable inventory from full wastage — on top of the 30% reduction achieved by right-sizing orders.
Implementation Timeline
Month 1 established the data pipeline foundation. The client used Increff WMS for warehouse management and a custom POS system across stores. We built bidirectional API integrations: real-time stock level reads from all 40 store POS systems (updated every 15 minutes), daily sales transaction exports at SKU × store × hour granularity, and supplier delivery confirmation data from the Increff WMS. Data quality validation revealed that 12% of SKUs had miscoded stock levels due to historical receiving errors. A one-time reconciliation exercise corrected these discrepancies; going forward, automated data quality checks flag suspicious stock level changes (sudden drops or spikes) for investigation before they enter the forecasting model.
Month 2 was the most analytically intensive phase: cleaning and enriching 3 years of daily sales data across 4,500 SKUs and 40 stores — 197 million data points — into a training-ready feature set. Cleaning tasks included: removing stockout days from the sales history (a day where sales were zero because of stockout, not zero demand, would train the model to predict zero demand incorrectly); identifying and flagging promotional periods in historical data; standardising SKU codes that had changed during system migrations; and imputing missing data for the 3 stores that had POS system outages during the 3-year period. Feature engineering created 47 predictive features per observation including lag features (sales 1, 7, 14, 28 days prior), rolling averages (7-day, 30-day), day-of-week and month-of-year encodings, holiday proximity features, weather features aligned by store geography, and promotional flags.
Month 3 trained and validated the forecasting models. LightGBM models were trained per product category (not per individual SKU — this prevents overfitting on low-velocity products while allowing the model to learn category-level demand patterns). Category models then apply SKU-specific velocity adjustments derived from the SKU's historical share within its category. Prophet time series models were trained for the 340 SKUs with strong seasonal patterns (fresh produce, festival foods, seasonal beverages). Model validation used a 90-day holdout period (the most recent 90 days of historical data were withheld from training and used as out-of-sample validation). Weighted Mean Absolute Percentage Error (WMAPE) — weighted by sales volume so high-velocity SKUs matter more than slow movers — reached 6.2%, corresponding to 93.8% weighted accuracy.
Month 4 built the production infrastructure that converts model outputs into daily operational actions. Apache Airflow orchestrates a 6-step daily pipeline: POS data ingestion and validation (2 AM), feature generation (3 AM), model inference for all 180,000 forecast points (4 AM), purchase order draft generation (5 AM), expiry risk computation for perishables (5:30 AM), and dashboard and notification delivery (6 AM). The buyer-facing dashboard was built as a React web application with mobile-responsive design — store managers use it on tablets and phones. Purchase orders can be approved, modified, or rejected with reason codes directly from the dashboard. Approved orders are automatically submitted to supplier EDI systems or emailed in the required format for suppliers without EDI capability.
Month 5 ran a 4-week parallel operation: buyers continued their existing Excel-based process while simultaneously receiving AI-generated purchase orders through the new dashboard. This parallel run served two purposes — buyer confidence building (watching the AI's recommendations prove accurate builds trust more effectively than any training session) and model calibration (systematic comparison of AI recommendations versus buyer decisions identified specific SKU categories where the model needed additional feature engineering). At the end of Month 5, buyer acceptance rate for AI-generated POs without modification reached 71%. Month 6 completed the transition to AI-led ordering, decommissioned the Excel-based process, and launched the full expiry risk dashboard for all 40 stores. A management reporting layer was added showing chain-level forecast accuracy, waste metrics by category, and stockout tracking against the pre-AI baseline.
Results
Perishable waste dropped from 18% to 12.6% of perishable inventory value monthly — a 30% relative improvement. In absolute terms, this represents ₹64.8 lakhs in annual waste savings (30% of ₹18L monthly perishable waste × 12 months). Markdown promotions triggered by the expiry risk dashboard saved an additional 4% from full wastage — these items were sold at 30–40% discount rather than being written off entirely. The biggest improvements were in dairy and fresh produce, where demand is most predictable from day-of-week and weather signals, and where the cost of error is highest.
Daily stockout events across all 40 stores dropped from 230 to 81 — a 65% reduction achieved within 3 months of full AI-led ordering. High-velocity SKUs (top 200 by sales volume) achieved a 99.2% in-stock rate — essentially eliminating stockouts on the products that matter most for customer satisfaction. For these key items, the improvement from 87% in-stock rate to 99.2% represents the difference between customers experiencing a stockout once every 8 visits versus once every 125 visits. The commercial impact extends beyond direct sales recovery: management estimates that the stockout reduction contributed to a measurable improvement in customer satisfaction scores and reduced the rate at which loyal customers cited "items not available" as a reason for reducing visit frequency.
Right-sizing inventory across all 40 stores — ordering precisely what is needed rather than a gut-feel buffer above it — reduced average inventory holding by ₹3.2 crore. This capital is now available for operational use: the chain used a portion to fund store renovation at two high-potential locations and reduce short-term borrowing costs. At the chain's cost of capital of 12% per annum, the ₹3.2 Cr working capital release generates ₹38.4 lakhs in annual financial benefit even before the waste savings are counted. The inventory optimisation also improved stock freshness scores measured by customer surveys — products are on shelves for shorter periods before sale, meaning fresher products reach customers more consistently.
Three buyers who previously spent 3 hours each daily on routine purchase order generation are now freed from that administrative burden entirely. Their time is redirected to activities that create competitive advantage: supplier negotiation (the chain negotiated improved credit terms with 4 major suppliers using forecast commitment data as leverage), private label development (two new private label categories were launched using the capacity freed from daily PO work), and promotional planning (structured analysis of promotional uplift data, now available from the AI system, enables more precise promotional investment decisions). This capacity reallocation represents a qualitative improvement that is difficult to quantify financially but is recognised by the management team as the most lasting organisational benefit of the project.
ROI Breakdown
The financial case for AI inventory forecasting in a multi-store retail chain is built on three independent value streams: waste reduction (direct cost saving), stockout recovery (revenue retention), and working capital release (financing cost reduction). Together, these generate a compelling first-year return that compounds in subsequent years as the model accumulates more training data and forecast accuracy improves further.
| Value Stream | Calculation | Annual Value |
|---|---|---|
| Perishable waste reduction (30%) | 30% × ₹18L/month × 12 months | ₹64.8 L |
| Markdown rescue savings (additional 4%) | Recovered at 35% average discount vs 100% write-off | ₹9.4 L |
| Stockout sales recovery | 65% reduction in ₹2.15 Cr annual lost sales | ₹1.40 Cr |
| Working capital cost saving | ₹3.2 Cr freed × 12% cost of capital | ₹38.4 L |
| Buyer time reallocation value | 3 buyers × 3 hrs/day × 250 days × ₹600/hr | ₹13.5 L |
| Total annual value | — | ₹3.15 Cr |
| System build + Year 1 operating cost | ₹35 lakhs | |
| Return on Investment — Year 1 | 9x ROI | |
Tech Stack
FAQ
Related Services
Get a free forecast accuracy analysis on your last 6 months of sales data. We'll show you what AI forecasting would have saved you.
Get Free Forecasting Analysis