Innovative AI Solutions | AI Development, Web & Mobile Apps – Delhi, India
FMCG · DEMAND FORECASTING · SUPPLY CHAIN AI

AI Demand Forecasting for FMCG — 35% Better Accuracy

We built a sales forecasting ML system for an FMCG company with 800 SKUs distributed through 5,000 outlets — improving forecast accuracy from 68% to 92%, reducing safety stock by 25%, and cutting stockouts at distributor level by 50%.

35%Better Forecast Accuracy
25%Safety Stock Reduced
50%Fewer Distributor Stockouts
₹5.2CrWorking Capital Freed
Build FMCG Supply Chain AI All Case Studies

Why FMCG Demand Forecasting Is Uniquely Difficult in India

India's FMCG market is one of the most complex demand forecasting environments in the world. The structural factors that make Indian FMCG demand uniquely difficult to predict are precisely what make AI forecasting most valuable.

🗓️

Festival Calendar Complexity

India has more than 40 culturally significant festivals that affect consumer spending patterns — not uniformly, but in hyper-local ways that require granular geographic modeling. Diwali drives a 60–80% demand spike for snack foods nationally, but the timing of this spike varies by geography (North India celebrates 5–7 days before South India buys gifting packs). Eid demand for certain food categories in UP and Bihar follows a lunar calendar that shifts by 10–12 days each year. Harvest festivals in Maharashtra, Punjab, and Tamil Nadu create regional peaks that do not appear in national aggregate data. A moving-average forecast treats all these events as "unexplained variance" to be buffered with safety stock. An AI forecast trained on the festival calendar and regional consumption patterns can anticipate each of these spikes 4–6 weeks in advance with sufficient accuracy to plan production and procurement without emergency orders.

🌦️

Weather Creates 30–40% Demand Swings

Temperature and monsoon patterns create dramatic demand swings in beverage, snack, and personal care categories that no calendar-based model can capture. A heatwave in May in Delhi increases cold beverage demand by 35–45% beyond the expected seasonal baseline. An early monsoon reduces summer beverage demand by 20–25% — in the same week a manufacturer had planned maximum production. The correlation between weather and demand is well-documented in FMCG research, but capturing it operationally requires integrating weather forecast APIs into the demand planning system and modeling the category-specific, geography-specific weather-demand relationship — work that is beyond the capability of Excel-based planning but well within the scope of an ML forecasting system. Incorporating 5-day weather forecasts improved the client's beverage category forecast accuracy by 18 percentage points alone.

📊

SKU Complexity: 800 Products × 12 States

An FMCG company with 800 SKUs distributed across 12 states is managing 9,600 distinct SKU-geography demand streams simultaneously. Human planners, even with sophisticated Excel models, cannot maintain accurate, updated forecasts for 9,600 streams without either significant simplification (aggregating SKUs and geographies in ways that mask the local demand variations that matter for stock allocation) or significant headcount (dedicated analysts for each category, each channel, and each region). ML forecasting systems handle 9,600 streams as naturally as 9 — the computational cost of modeling one more stream is negligible. This scalability advantage is where AI creates the most durable and defensible value in FMCG planning: the human planning system's accuracy degrades as complexity grows; the AI system's accuracy improves as it accumulates more data from more streams.

Excel Forecasts Missing Seasonal Spikes by 40%

The FMCG company (snacks and beverages) forecasted production using 3-month moving averages in Excel. This approach completely failed during key periods: summer (40% volume spike for beverages), Diwali (snacks demand doubles in North India), and monsoon (regional demand shift patterns). Production planning lagged reality by 4–6 weeks, resulting in simultaneous overproduction of slow SKUs and stockouts of fast-moving products. The supply chain team used the phrase "firefighting" so frequently in their meeting vocabulary that it had stopped being a metaphor — it was their actual daily activity.


Distributors were complaining of stockouts on 15–20% of orders. The trade marketing team was running promotions without coordinating with production, creating demand spikes that supply couldn't meet. A 20%-off promotion on a beverage brand in April — timed to a summer peak — generated a demand spike that the production team had no visibility on, leading to a 3-week stockout at 35% of distributors. The promotion generated ₹1.8 Cr in incremental sell-in; the stockout cost an estimated ₹4.2 Cr in lost sales and retailer relationship damage. This disconnect between trade marketing, sales, and supply chain was systemic — not a one-off failure but a quarterly pattern.


The S&OP process was a monthly meeting where everyone disagreed and nothing was resolved. Sales presented optimistic forecasts biased by sales targets. Supply chain presented conservative forecasts biased by fear of overproduction. Finance used the previous quarter's actuals with a standard growth factor. All three numbers were different; no one trusted any of them; and production planning proceeded on the supply chain team's conservative estimate, systematically under-planning relative to actual demand. The missing ingredient was not process discipline but a shared, data-driven number that all functions could see was based on evidence rather than departmental bias.

AI That Sees Demand 4 Weeks Before It Arrives

A multi-factor demand forecasting system that incorporates seasonality, promotions, distributor secondary sales, weather, and competitive pricing signals.

01

Granular SKU + Region Forecasting

The core forecasting engine uses a LightGBM + LSTM ensemble — LightGBM for capturing the complex feature interactions that drive demand (SKU × season × geography × promotion × weather), LSTM for capturing temporal patterns (week-of-year seasonality, trend momentum, lagged demand responses to promotions). The ensemble was chosen after comparative evaluation against Facebook Prophet, ARIMA, and XGBoost on the client's historical data: the LightGBM+LSTM ensemble produced 23% lower MAPE than the best single-model approach. Forecasts are generated for each of 800 SKUs across 12 states at weekly granularity, 4 weeks ahead. Updated every Monday morning with the latest secondary sales data from the distributor management system. The model forecasts at state level (because production and primary distribution happen at state level) but also generates regional sub-forecasts (metro vs. tier-2 vs. rural) for SKUs where the company's data shows significantly different demand patterns across outlet types.

02

External Signal Incorporation

The forecasting feature set was expanded beyond historical sales data to include external signals that predict demand before it appears in sales data: (1) Weather forecasts from the India Meteorological Department API — integrated for beverage and seasonal snack categories with a statistically validated demand-response model calibrated on 3 years of weather-sales correlation data. A 3°C temperature increase above seasonal average in any state predicts a 12% uplift in cold beverage category demand in that state for the following week. (2) Festival calendar — encoded as feature variables for each culturally significant event with geographic scope flags: national festivals (Diwali, Eid, Holi), regional festivals (Pongal, Durga Puja, Baisakhi), and harvest-related consumption events that affect rural distribution significantly. (3) School calendar — term-time vs. school holiday patterns significantly affect snack categories in segments with high child consumer purchase influence. (4) Competitor pricing signals — obtained through distributor network intelligence, where distributor sales reps report competitor trade promotions that they observe in the field. This last source is manual and imperfect, but even partial competitor intelligence materially improves the forecast for categories where price-switching is significant.

03

Promotion Impact Modeling

Marketing team inputs planned promotions (BOGO, price cuts, display schemes, BTL activations) into the system via a promotion management module 3–4 weeks before the promotion goes live. Each promotion is characterized by: type (price reduction %, BOGO, bundle, display scheme), geography (national, zone, state, or district), SKU or brand scope, duration in weeks, and channel (general trade, modern trade, e-commerce). The model estimates the promotion uplift based on historical promotion response rates — calibrated on 3 years of promotion data where outcomes were measured against baseline. A 15% price reduction on the beverage brand in Maharashtra in summer produces a 28–32% volume uplift based on historical response curves; the same promotion in winter produces only 8–12% uplift because the price elasticity is lower when the weather removes the intrinsic demand driver. Production planning sees the model's demand spike prediction for the promotion 3–4 weeks before launch — sufficient lead time to plan production and distributor stock build-up without emergency orders. In the 12 months post-deployment, no promotion-related stockout has occurred — a perfect record against the pre-deployment baseline of 4–6 promotion-related stockouts annually.

04

S&OP Collaboration Platform

All forecasts are visible to sales, marketing, supply chain, and finance in one platform — replacing the monthly disagreement meeting with a data-driven weekly sync. The S&OP platform shows: the AI consensus forecast with confidence intervals (for production planning, the planner needs to know not just the point estimate but the range of outcomes they should plan for); the previous forecast vs. actual comparison (for model trust-building and bias tracking); promotion impact overlays (showing the baseline forecast and the promotion-adjusted forecast separately, so the promotion uplift is visible as a distinct, trackable component); and a manual override module where sales or supply chain can adjust the AI forecast with a mandatory reason code. Overrides are tracked and their accuracy measured after the fact — producing a bias analysis report that shows which function's manual adjustments improved versus degraded forecast accuracy. This feedback mechanism has a predictable effect: over 6 months, the frequency of manual overrides declined 62% as each function accumulated evidence of which of their instincts the AI model correctly captured and which ones added noise.

16-Week Build from Data Audit to Live S&OP

A structured rollout through five phases — from data infrastructure setup through live S&OP meeting integration with all business functions.

W1-3

Data Infrastructure & Secondary Sales Integration

The first decision in every FMCG forecasting project is which sales data to forecast: primary sales (what the manufacturer sells to distributors) or secondary sales (what distributors sell to retailers). Primary sales data is readily available from the ERP system but reflects distributor stocking behavior, not consumer demand — a distributor loading up inventory ahead of a promotion inflates primary sales and creates a "pipeline fill" effect that makes the underlying demand trend invisible. We built a secondary sales data ingestion pipeline from the client's DMS (Distributor Management System) — extracting weekly SKU-level sales from distributors to retail outlets across all 5,000 outlet relationships. DMS data required significant cleaning: 12% of records had duplicate entries from connectivity sync errors; 8% had SKU coding inconsistencies across distributor systems; and 3% of outlets had implausibly large single-week volumes indicating data entry errors. Cleaned secondary sales data was loaded into a PostgreSQL data warehouse, validated against the corresponding primary sales volumes for reasonableness, and made available to the forecasting pipeline.

W4-6

Feature Engineering & External Data Integration

Feature engineering was the phase where domain knowledge from the FMCG client was most valuable. We worked with the category management and trade marketing teams over 3 weeks to enumerate every factor that they intuitively knew affected demand but had never been able to quantify. This process produced a feature list of 64 candidate variables: seasonal indices by month, temperature deviation from historical average, rainfall deviation, festival dummy variables with lead/lag windows, school calendar indicators, promotion type and intensity variables, distributor coverage change indicators, and competitor promotion binary flags. For each variable, we ran correlation and information gain analysis against the historical demand data to determine which features had genuine predictive power versus which were believed to matter but showed no statistical evidence of influence. The 64 candidate features reduced to 38 features with validated predictive power — the others were noted as "understood by the team but not measurable in data" and flagged for potential future data collection.

W7-10

Model Training, Validation & Accuracy Benchmarking

Model training used 3 years of weekly secondary sales data across 800 SKUs and 12 states. The train-validation split used a temporal approach (not random splitting): the model was trained on the first 30 months and validated on the final 6 months — mimicking the real-world scenario where the model is evaluated on future data it has never seen. This temporal split is critical for FMCG forecasting because random splits would allow the model to "cheat" by seeing data from the same month across different years, which would overstate accuracy on the seasonal patterns. Baseline comparison was run against the client's existing Excel moving-average model and against three ML alternatives (ARIMA, Prophet, XGBoost). The LightGBM+LSTM ensemble achieved 92% forecast accuracy (MAPE of 8%) on the held-out validation period — versus 68% for the existing Excel model (MAPE of 32%). Improvement was consistent across all product categories and geographies, with the largest improvement in the beverage category during summer months — exactly the period where the existing model was most wrong and where forecasting error had the largest operational cost.

W11-13

S&OP Dashboard & Production Planning Integration

The S&OP platform was built as a React web application accessible to all functions — unusual in a company where supply chain used SAP, sales used an internal CRM, marketing used Excel, and finance used Tally. A platform that required any function to change their primary tool would have faced adoption resistance. Instead, the S&OP platform was designed as a read-only reporting and collaboration layer that pulls data from all function-specific systems rather than replacing them. Forecasts flow into SAP's production planning module via integration using SAP's Plant Maintenance RFC interfaces — production planners continue working in SAP but now receive AI-generated demand inputs rather than Excel-uploaded planner estimates. The promotion management module was integrated with the marketing team's promotion planning calendar in Google Sheets (the system they were already using) via Google Sheets API — promotion plans entered in Sheets automatically appear in the S&OP platform and trigger the promotion impact modeling run.

W14-16

Shadow Running, S&OP Process Change & Full Go-Live

For 3 weeks, the AI forecast ran in shadow mode alongside the existing Excel process. The weekly S&OP meeting still used the Excel numbers for actual decisions, but the AI forecast was presented on the same screen. This side-by-side comparison was the most powerful adoption tool: in week 1 of shadow running, the AI forecast predicted a major stockout on the flagship biscuit SKU in Maharashtra 3 weeks ahead — a prediction that the Excel model showed nothing about. The supply chain team ordered a production build-up. The stockout did materialize in week 3, but they had already built the buffer and fulfilled all distributor orders. This single incident — where the AI was visibly right and the Excel model was visibly wrong, with real stakes — created more adoption buy-in than any training session could have. By week 16, all five functions had agreed to use the AI forecast as the single company number going into the S&OP meeting, with manual override reserved for specific situations with explicit justification.

Supply Chain Performance Transformed

🎯

68% → 92% Forecast Accuracy

MAPE (Mean Absolute Percentage Error) improved from 32% to 8% at the SKU-month level. Summer and festival season spikes now predicted accurately 4 weeks in advance. The improvement is most dramatic in the categories that were most difficult to forecast manually: beverages in summer (from 45% error to 7% error), snacks during Diwali (from 52% error to 11% error), and rural distribution during harvest seasons (from 38% error to 9% error). These were the exact categories where the company was experiencing its most painful stockouts and overproduction cycles.

📦

25% Less Safety Stock

With better forecasts, safety stock buffers were reduced — freeing ₹5.2 Cr in working capital previously tied up in "just in case" inventory. The safety stock reduction was implemented gradually: as the AI forecast demonstrated accuracy over the first 3 months, safety stock parameters were recalibrated category by category to reflect the new, lower forecast error. The ₹5.2 Cr freed from safety stock is now available for trade marketing investment, new product launches, and debt reduction. CFO described this as "the first time in 8 years we've been able to reduce inventory without either accepting higher stockout risk or compromising service levels."

🚛

50% Fewer Distributor Stockouts

Distributor fill rate improved from 83% to 96%. Distributor complaints about stockouts dropped from top complaint to #6 in quarterly reviews. Improved distributor relationships have produced secondary benefits: 3 distributors who had been considering dropping the company's brand in favor of a competitor's restocked their range and increased their investment in display and promotion following the service level improvement. One zone sales manager cited the fill rate improvement as the single biggest factor in a 14% sales volume increase in their zone, because distributors who trust supply consistency invest more in pushing the brand to retailers.

🤝

S&OP Finally Works

Sales, marketing, and supply chain teams now share one number. Decisions on production, procurement, and promotions made from same data. Alignment improved across functions. The monthly S&OP meeting has been restructured from a 3-hour debate over whose forecast number is right to a 45-minute review of the AI forecast, exception management for SKUs where manual judgment is being applied, and promotion planning coordination. The supply chain director described the change: "We used to leave the meeting agreeing to disagree. Now we leave the meeting with one plan that everyone is aligned on — because the AI gives us a neutral starting point that no one owns politically."

The Financial Case for AI Demand Forecasting

Quantified business impact for an FMCG company with 800 SKUs, ₹220 Cr annual revenue, and distribution through 5,000 outlets — covering working capital, service level, and revenue impact.

Value ComponentCalculation BasisAnnual Value
Working Capital Release (Safety Stock)₹5.2 Cr freed at 12% cost of capital — annual capital cost saving₹62,40,000
Stockout Revenue Recovery50% fewer stockouts × estimated ₹3.8 Cr annual lost-sale impact₹1,90,00,000
Overproduction Waste ReductionReduced slow-SKU overproduction → fewer write-offs and markdown sales₹45,00,000
Promotion Planning EfficiencyZero promotion-related stockouts (previously 4–6/year at ₹50L impact each)₹2,00,00,000
Emergency Procurement Premium EliminatedRush orders to suppliers eliminated — previously ₹18L/year premium₹18,00,000
System Build + Year 1 InfrastructureDevelopment + DMS integration + SAP integration + S&OP platform + maintenance-₹32,00,000
Net Year 1 Return₹4,83,40,000

Technologies Used

PythonLightGBMLSTM (TensorFlow)Facebook ProphetApache AirflowFastAPISAP IntegrationDMS (Distributor Management)PostgreSQLReact S&OP Dashboard

About This Project

How do you handle SKUs with very intermittent demand? +
For slow-moving and intermittent SKUs (demand in only 20-30% of periods), standard time-series models fail. We use Croston's method and its variants for intermittent demand forecasting. For near-zero velocity SKUs, we switch to minimum order quantity replenishment rules rather than probabilistic forecasting. SKUs are automatically classified by velocity tier (A, B, C, D) and the appropriate modeling approach is applied per tier — no manual classification required. This tiering is reviewed quarterly as SKU velocities change (new product launches initially appear as slow-movers before building velocity; declining SKUs move from fast to slow). The tiering system ensures that the forecasting model's complexity matches the data richness available for each SKU.
What secondary sales data does the model need? +
We use distributor secondary sales data (what distributors sold to retailers) — not just primary sales to distributors. This removes the "pipeline fill" distortion that makes primary sales data unreliable for demand forecasting. We ingest data from the client's DMS system and supplement with outlet-level sell-through from their field sales app. Minimum data requirement: 18 months of weekly secondary sales by SKU and state. With less data, the model can work but will have reduced accuracy for seasonal patterns (it needs to see at least one full seasonal cycle to learn the pattern). Companies without a DMS can start with primary sales data while implementing a DMS — we can build the model on primary data and migrate to secondary data as the DMS comes online, typically within 3–4 months.
Can the model be used for production planning and procurement too? +
Yes — and we built the entire chain. Demand forecast → Net requirement calculation (after inventory netting) → Production plan (considering capacity constraints) → Raw material procurement plan (considering lead times and MOQs). The system generates a weekly production plan and 4-week procurement plan automatically — ready for planner review and override. For clients on SAP, the production plan is pushed directly to SAP PP as planned orders; planners review and release in SAP as usual, but with AI-generated demand inputs rather than manually estimated ones. Raw material procurement recommendations are generated with supplier-specific lead time and MOQ constraints, enabling the procurement team to place orders based on the AI's forward-looking demand plan rather than reacting to production emergencies.
How does the model handle new product launches (NPDs) with no historical data? +
New product launches are the most challenging forecasting scenario because they have zero historical sales data. We address this through: (1) Analog product matching — identifying 3–5 "analog" products from the company's portfolio or publicly available market data that are most similar to the NPD in terms of category, price point, pack size, and channel. The analog products' launch trajectories are used to build a launch curve prior for the NPD. (2) Market research integration — if consumer research or trade sampling results are available pre-launch, we calibrate the launch curve based on test market data or distribution build targets. (3) Conservative safety stock: NPD forecasts carry deliberately higher safety stock coverage (10–12 weeks vs. the standard 4–6 weeks for established SKUs) until 3 months of actual sales data allow the model to transition from analog-based to data-driven forecasting. This approach has produced NPD forecast accuracy of 65–75% in the first 90 days — significantly better than the 40–50% typical of manual NPD forecasting.
How do we handle the transition from Excel-based S&OP to AI-driven S&OP without creating resistance? +
Change management for S&OP process change is arguably more important than the technical implementation. We consistently recommend a shadow-running approach: run the AI forecast alongside the existing process for 4–8 weeks without requiring anyone to use it for actual decisions. Let each function see their own historical forecast accuracy versus the AI's accuracy on the same historical data — this data-driven comparison is more persuasive than any training session. The functions most resistant to AI forecasting are typically sales (who fear losing the ability to influence the demand number upward to protect their targets) and supply chain (who fear reducing safety stock based on model accuracy that they don't yet trust). The shadow period addresses both concerns empirically. A deliberate override tracking and feedback loop is also critical: if functions know their overrides will be tracked and their accuracy measured after the fact, they apply overrides more judiciously — and the quality of human judgment layer improves naturally over time.
How often does the model retrain, and does accuracy improve over time? +
The model retrains weekly on new secondary sales data using Apache Airflow — every Monday morning, the previous week's DMS data is ingested, and the model is retrained on the full rolling 3-year window of data. This weekly retraining ensures the model learns from recent demand patterns and does not rely solely on older data that may not reflect current market conditions (new competitors, pricing changes, distribution expansion). Accuracy does improve over time in two ways: (1) More data improves the model's statistical confidence, particularly for low-volume SKUs and for unusual scenarios (a specific festival pattern seen only once in 18 months of training data becomes much more confidently modeled after seeing it 3 times in 36 months); (2) The correction feedback from manual overrides and post-hoc accuracy analysis is incorporated monthly — any systematic bias in the model (consistently under-forecasting in Maharashtra, for example) is identified and corrected in the feature engineering layer. Client's MAPE has declined from 8% at launch to 6.2% at 12 months — a 23% further improvement on top of the initial 35% gain from switching from Excel to AI.

Services Used in This Project

Replace Excel Forecasts with AI-Driven Demand Planning

Get a free forecast accuracy benchmark on your historical sales data — we'll show you the improvement AI can deliver in 2 weeks.

Get Free Forecasting Assessment