AI demand forecasting for spare parts and inventory optimization

After-sales vehicle service faces a paradox: hundreds of thousands of SKUs, of which 80% sell less than once a month, yet shortage of a single critical part stops a car for a week. Demand for new parts (cold start) is virtually impossible to predict with classical methods. We solve this by combining ABC classification with ML models adapted to different demand types. Our approach reduces inventory levels by 15–25% while maintaining or improving service levels — proven across 50+ projects in parts distribution.

Why Croston and TSB are necessary for intermittent demand

Most parts sell infrequently and irregularly. ARIMA or Prophet on such demand yield MAPE >100%. Croston separately forecasts the interval between sales and transaction size, while TSB (Teunter-Syntetos-Babai) adds adaptation for obsolescent parts — reducing bias by 30–50%. On real distributor data, Croston achieved MAPE 78% vs 142% for ARIMA — 1.8 times more accurate.

from statsforecast.models import CrostonOptimized, IMAPA, TSB

models = [
    CrostonOptimized(),
    IMAPA(),
    TSB(alpha_d=0.1, alpha_p=0.1)
]

Comparison of forecast methods for intermittent demand:

How park-based forecast improves accuracy

Parts demand is driven not only by sales history but by the vehicle fleet. We build a model: demand = number of cars × failure probability by age. Data source — registration databases (fleet by model and year).

def park_based_forecast(part_number, region):
    applicable_models = parts_catalog.get_applicable_models(part_number)
    park = vehicle_registration_db.count(
        models=applicable_models,
        region=region,
        age_range=(2, 20)
    )
    failure_curve = get_failure_rate_curve(part_number)
    expected_demand = sum(
        park[age] * failure_curve[age]
        for age in range(2, 21)
    ) / 12
    return expected_demand

This method is especially effective for new parts or product launches (cold start) when no sales history exists.

Safety stock optimization considering cost asymmetry

Classic newsvendor model: we optimize safety stock not by symmetric normal distribution but by real costs of stockout (customer downtime, lost loyalty) and holding. The cost of a critical part shortage can be 50 times the holding cost — safety stock for such SKUs is raised to z=3.0 vs standard z=1.65. Typical savings from this approach reach 2 million RUB per year for an average warehouse.

from scipy.stats import norm

def optimal_safety_stock(mean_demand, std_demand, lead_time_days,
                          service_level=0.95, holding_cost_rate=0.25,
                          stockout_cost=50.0):
    cu = stockout_cost
    co = holding_cost_rate * unit_cost / 365
    critical_ratio = cu / (cu + co)
    z = norm.ppf(critical_ratio)
    demand_during_lt = mean_demand * lead_time_days
    std_during_lt = std_demand * np.sqrt(lead_time_days)
    safety_stock = z * std_during_lt
    reorder_point = demand_during_lt + safety_stock
    return safety_stock, reorder_point

How ML Ops keeps models in production

Models degrade over time: vehicle fleet, seasonality, assortment change. We set up an ML Ops pipeline: automatic retraining when accuracy drops (alert at MAPE > 90%), metric logging in MLflow, A/B testing of new versions. This ensures stable forecast performance without monthly manual intervention.

Managing obsolescent parts (phase-out)

A discontinued model — demand declines but not instantly. TSB model detects the downward trend, and our script identifies EOL signs:

def detect_obsolescence_risk(part_number, sales_history):
    trend = np.polyfit(range(len(sales_history)), sales_history, 1)[0]
    park_decline = get_park_trend(part_number)
    if trend < -0.1 and park_decline < -0.05:
        return 'phase_out', suggest_final_buy_quantity(part_number)
    return 'active', None

Final buy — optimal order quantity for 5–7 years of warranty service. We account for the fact that after EOL parts become more expensive or unavailable.

Multi-echelon: reducing the bullwhip effect

In the OEM → distributor → dealer chain, demand variation amplifies. We synchronize safety stock across all levels: the higher the dealer buffer, the lower the distributor safety stock. Vendor Managed Inventory (VMI) replenishes the dealer's warehouse in real time — reducing overall chain inventory by 15–20%.

What is included in the work

1. Data audit: sales history, vehicle fleet, parts catalog.
2. Building ML models: Croston, TSB, park-based, seasonal factors.
3. Safety stock and reorder point optimization (differentiated service levels).
4. Phase-out management and final buy recommendations.
5. Dashboard with forecasts, stockout/overstock alerting.
6. Integration with WMS/ERP, VMI setup.
7. Team training and operational documentation.

Payback period is less than 12 months. Contact us — we will assess your inventory portfolio and show the potential for inventory reduction. Our team: 10+ years in ML solutions for logistics, 50+ implemented projects, certified specialists in Croston method.

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.