Development of AI System for Demand Forecasting

Manual demand forecasting in Excel leads to MAPE of 20–30%. For a retailer with 50,000 SKUs, this translates to millions of rubles in losses monthly due to write-offs or stockouts. We automate ML pipelines that reduce error to 8–12% through promo modeling, hierarchical reconciliation, and automatic feature engineering. Regardless of vertical — manufacturing, FMCG, e-commerce, or services — the key problems are the same: data quality, promo effects, and scaling. Our solutions are adapted to business specifics: for retail — short horizons and promo lifts, for manufacturing — long-term forecasts and macro indicators. Reducing MAPE from 20% to 8% cuts write-offs by 15–20 million rubles per year for a chain of 500 stores. Savings on logistics costs reach 10–12 million rubles. A common mistake is ignoring the promo calendar and seasonal patterns, leading to skewed forecasts. We use automatic feature selection and model ensembles to increase robustness.

Vertical-Specific Challenges

Manufacturing

Horizon 3–6 months due to lead time, key KPI is accuracy for raw material planning. Data: historical orders, macro indicators.

FMCG / Retail

Horizon 1–4 weeks, promo lifts up to +300%. Separating baseline demand and incremental is the main difficulty.

E-commerce

Horizon 1–7 days, SKU-level, extreme seasonality (Black Friday). Global models for 100,000+ SKUs are required.

Services

No physical inventory, but capacity (call center operators, servers). Load is forecasted, not goods.

How ML Models Handle Promo Effects?

Promotions are the largest source of error. Decomposing Total Demand = Baseline + Incremental Lift allows modeling the lift separately. Example: a LightGBM regressor for lift takes features like discount_pct, mechanic, display_flag, brand_strength and outputs a coefficient — e.g., 1.85 (+85%). Cross-SKU effects (cannibalization and halo) adjust forecasts across the category via a cross-elasticities matrix. LightGBM outperforms Prophet by 1.15–1.2 times in MAPE for short-term forecasts with promotions.

lift_features = {
    'discount_pct': 20.0,
    'mechanic': '2+1',
    'display_flag': 1,
    'leaflet_flag': 0,
    'competitor_promo': 0,
    'category': 'soft_drinks',
    'brand_strength': 0.8,
    'seasonality_index': 1.2
}
predicted_lift = lift_model.predict([lift_features])

Why Hierarchical Forecasting Is Critical for 10,000 SKUs?

Hierarchy: Total → Category → Brand → SKU → Location. Manual reconciliation of 500,000 forecasts daily is impossible. MinT (Minimum Trace) provides unbiased estimates at all levels. Comparison of methods:

We use a hybrid: statistical methods for long horizons and ML for short, aggregated via MinT.

New Product Introduction (NPI)

New SKUs with no history are a common pain. Three approaches:

• Analog-based: forecast based on sales of similar products at launch.
• Attribute-based: regression on characteristics (brand, category, price).
• Bayesian prior: initial forecast = analog, updated with first sales.

Forecasting System Architecture

Data Sources → Feature Engineering → Model Training → Forecast → Activation

Data Sources:
├── Internal: ERP sales, WMS, CRM
├── External: macro data, weather, search trends
└── Promotional: trade calendar, planned campaigns

Feature Engineering (dbt / Spark):
├── Temporal lags: t-1, t-7, t-28, t-52 (weeks)
├── Rolling aggregations: 4w, 13w, 52w
├── Promotional features: lift estimation, channel flags
└── External features: weather index, macro indicators

Model Training (MLflow):
├── Baseline: Seasonal Naive, ETS
├── Statistical: Prophet, SARIMA
├── ML: LightGBM, DeepAR
└── Ensemble: Stacking / Weighted Average

Forecast Generation:
└── Hierarchical reconciliation → SKU × Location prognoses

How to Build a Baseline Forecast for New SKUs?

1. Collect data on launches of similar SKUs over the last 2 years.
2. Extract attributes: brand, category, price segment, launch season.
3. Build a regression model to predict the first 4 weeks of sales.
4. Use Bayesian update: adjust the forecast weekly based on actual sales.

This approach gives start accuracy WMAPE 15–18% vs. 30% for naive average.

Forecast Integration and Activation

The system exports forecasts to S&OP (SAP IBP, Anaplan) via API and generates purchase orders for VMI. Accuracy tracking: 1 – WMAPE on a dashboard. We guarantee transparency: model card, SHAP reports, pipeline documentation. Thanks to forecast accuracy, our clients reduce storage costs by 20–30%.

What's Included in the Work (Deliverables)

• Audit of data sources and cleaning.
• Feature engineering pipeline (dbt/Spark).
• Model training and validation (MLflow).
• Architectural documentation and model card.
• Integration with ERP/S&OP via API.
• Team training and 3 months of support.

We will assess your project in 2–3 days — get in touch with us. 5+ years of experience, 30+ implemented demand forecasting systems for retail and manufacturing. Typical timelines: basic system with LightGBM for 1000+ SKUs — 6–8 weeks, full hierarchical system with NPI and reconciliation — 4–6 months. Get a consultation to discuss the details.

When does a time series forecasting model fail in production?

The CFO requests a quarterly sales forecast. An analyst builds SARIMA on three years of data, achieves MAPE 8.3% on the test set, and deploys. Two months later, the metric in production jumps to 23%. The root cause: the model was trained on pre‑COVID data, tested on a stable period, but production hit a promotion and supply chain disruption. Data leakage plus distribution shift—perfect notebook numbers, a broken forecast in reality. We have seen this pattern dozens of times across retail, fintech, and IoT. Our team has delivered more than 50 forecasting projects over 5+ years.

Incorrect cross-validation. Standard train_test_split for time series creates data leakage: the model sees future values during training. The correct approach is TimeSeriesSplit or walk‑forward validation with an expanding window.

Multiple seasonality. Hourly electricity consumption has three seasonalities: daily (24h), weekly (168h), yearly (8760h). SARIMA handles only one. Prophet can handle multiple but scales poorly to thousands of series.

Missing values and anomalies. A missing sensor reading is information (the sensor turned off), not NaN. Linear interpolation destroys this signal. Proper handling depends on the missingness mechanism.

Cold start. A new SKU in a 50,000‑item assortment has no history, yet a forecast is needed. Standard approaches fail; cross‑learning or feature‑based methods are required.

Why is model selection critical for your data?

Prophet (Meta) – a solid start for business data with clear seasonality and holidays. Fast setup, interpretable, built‑in outlier detection. Fails on irregular patterns and does not scale beyond ~10k series without parallelization.

Gradient boosting on features (LightGBM, XGBoost) – often underestimated. Engineer lags (t‑1, t‑7, t‑28), rolling means, day‑of‑week, holidays. The model trains on all series simultaneously, solving cold start via transfer learning. MAPE in retail often beats neural nets with proper feature engineering.

TFT (Temporal Fusion Transformer) – a transformer designed for interpretable forecasting with covariates. Built‑in variable selection, temporal attention, quantile outputs. Available in pytorch‑forecasting. Requires ~10,000+ records per series for stable training.

PatchTST – splits the series into patches (like ViT for images), capturing local patterns better than classic transformers. Excellent for long‑horizon forecasting (96–720 steps ahead).

N‑HiTS, N‑BEATS – attention‑free neural architectures, faster than TFT, competitive accuracy. N‑BEATS won the M4/M5 benchmarks for tasks without covariates.

How do we deploy TFT in production?

A typical pipeline via pytorch‑forecasting:

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="sales",
    group_ids=["store", "sku"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,  # 120 days
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,  # 28 days
    static_categoricals=["store_type", "category"],
    time_varying_known_reals=["price", "promo_flag"],
    time_varying_unknown_reals=["sales"],
    target_normalizer=GroupNormalizer(groups=["store", "sku"], transformation="softplus"),
)

A common mistake: the default target_normalizer (StandardScaler) breaks predictions for series with zero values (no sales on weekends). GroupNormalizer with transformation="softplus" is the correct choice for count data.

Case study: retail demand forecasting

A chain of 120 stores, 8,000 SKUs, 28‑day forecast horizon. The original system: SARIMA per series, MAPE 18.4%, retraining cycle – 6 hours. We replaced it with TFT on PyTorch + pytorch‑forecasting: a single model for all series, MAPE 11.2%, retraining – 40 minutes on an A10G. Feature importance via variable selection revealed that day_before_holiday influences more than the holiday date itself. Annual savings on inference alone exceeded $50,000.

Step‑by‑step configuration

1. Data collection and preparation. Handle missing values (mark NaN, interpolate only for technical failures), aggregate to required frequency, engineer covariates (holidays, promotions, prices).
2. Create TimeSeriesDataSet. Set group_ids (store + SKU), time index, forecast horizon. Choose target_normalizer based on target distribution.
3. Train a baseline. Prophet or LightGBM first – to understand complexity.
4. Train TFT. Use TemporalFusionTransformer with loss=QuantileLoss(), tune learning rate and hidden layer sizes.
5. Validate and interpret. Walk‑forward test, analyze variable selection, build attention heatmaps.

How to properly evaluate forecast quality?

RMSE alone is misleading – it over‑penalizes large values. Our standard set:

• MAPE – interpretable, unstable near zero.
• sMAPE – symmetric, avoids division by small numbers.
• MASE (Mean Absolute Scaled Error) – normalized relative to a naive seasonal forecast, ideal for comparing series of different scales.
• Pinball loss – for probabilistic forecasting, inventory management.

We guarantee a model card with these metrics on the validation set and walk‑forward results on at least 6 months of history.

What deliverables do you receive?

• Documentation of chosen architecture and hyperparameter rationale.
• Reproducible training and inference pipeline (Docker + CI/CD + Airflow/Prefect).
• Committed code with unit tests for key components.
• Team training: retraining, output interpretation, deployment of new versions.
• 3 months of post‑delivery support (consultations, bug fixes, fine‑tuning).

The model is deployed via FastAPI or Triton Inference Server. Retraining is scheduled (e.g., weekly) via Airflow with drift validation and automatic rollback if metrics deteriorate.

Process and timeline

We start with EDA: visualization, ADF test, STL decomposition, analysis of missing values and outliers. This takes 2–3 days but often reveals systemic data issues that block forecasting. Then we build a baseline (naive seasonal, Prophet), engineer features for LightGBM, and select a neural architecture if needed. Walk‑forward validation with a realistic horizon. Deployment via API with automatic retraining scheduled via Airflow or Prefect.

Timeline: MVP forecast on one data type – 3–6 weeks. Hierarchical forecasting system with automation – 2–5 months. Cost is calculated individually based on data volume, number of series, and required accuracy.

Our team consists of certified ML engineers (AWS ML Specialty, GCP Professional ML Engineer) with 5+ years on the market and over 50 completed forecasting projects. Contact us for a free analysis of your data – we will assess the task and provide initial recommendations within 1–2 days. Request a consultation to ensure your forecasts work in production, not just in a notebook.