AI-Powered ETA Prediction System for Last Mile and Long Haul Logistics

You have 200 couriers — each making 15–25 deliveries per day. A third of clients call the call center: "Where is my order?" The standard ETA of "by 18:00" gives a spread of ±2–3 hours. When a driver is an hour late, the client gets nervous. Especially if the order is dinner groceries? The result — cancellations, lower NPS, losses. We build ML systems that predict arrival time with 15–30 minute accuracy, using LightGBM, LSTM, and real-time data. Our cumulative experience — 20+ projects for logistics operators. We guarantee stable model performance in production and post-deployment support.

We have already implemented such projects for 20+ logistics operators in Russia and the CIS, accumulating over 5 years of ML experience in logistics. According to Uber Movement, LightGBM outperforms linear regression by 2–3 times in accuracy on historical data. Reducing call center load by 30–40% saves up to 3 million rubles per year. Contact us — we will audit your data in 2 days.

Why traditional ETA methods don't work

Linear regression based on average speed and distance does not account for:

• Traffic jams: during peak hours, speed drops by 2–3 times.
• Weather: rain or snow adds 15–40% time.
• Operational delays: queue at loading, time at stops.
• Historical patterns: routes with regular delays on specific days.

Traditional methods yield MAPE of 25–40%. An ML model reduces MAPE to 10–15%, saving up to 20% in logistics costs. For an operator with a fleet of 200 vehicles, savings from reduced failed deliveries can reach 5–7 million rubles per year.

How AI improves ETA accuracy

Feature engineering is key. We collect features from several sources:

Route data:

• Route distance (Google Maps / HERE / OpenStreetMap OSRM)
• Historical speeds on roads at different times of day
• Geofencing of pickup and delivery points

Operational data:

• Warehouse processing time (pick-pack-ship)
• Current queue at loading/unloading
• Number of stops en route to the target point

External factors:

• Weather: rain/snow/fog increase time by 15–40%
• Traffic events: accidents, roadworks, closures (TomTom TrafficStats, HERE Traffic API)
• Time patterns: morning peak 08–10, evening peak 17–19

Model architecture

Task: regression — predict time from dispatch to delivery in minutes.

Feature matrix:

features = {
    # Route
    'distance_km': route_distance,
    'n_stops': stops_remaining,
    'route_complexity': turns_per_km,

    # Time
    'hour_of_day': departure_hour,
    'day_of_week': departure_dow,
    'is_holiday': holiday_flag,
    'month': departure_month,

    # Traffic
    'historical_avg_speed': avg_speed_for_route_time,
    'current_traffic_index': live_traffic_score,  # 1.0 = normal, 2.0 = jam
    'weather_delay_factor': weather_impact_estimate,

    # Operational
    'shipment_weight_kg': weight,
    'vehicle_type': truck_van_bike,
    'driver_experience_days': driver_tenure,

    # Historical for this route
    'route_historical_eta': past_mean_eta_for_route,
    'route_eta_std': past_std_eta_for_route
}

Models:

• LightGBM Regressor: primary model for tabular data.
• Quantile Regression (p10/p50/p90): for ETA with confidence intervals.
• LSTM: if a sequence of intermediate GPS points is available.

Model comparison for ETA

Real-time ETA update

A static forecast at dispatch is not enough. The ETA must update dynamically:

Update triggers:

• Courier GPS tracking every 30 seconds.
• Traffic jam detected on route (traffic API polling every 5 min).
• Delay at previous delivery point.
• Weather event.

Online learning vs. static model: In production: the static model is retrained daily on new data. Real-time corrections via a kinematic motion model (speed + distance → updated ETA) without restarting the ML model.

def update_eta_realtime(current_position, destination, remaining_stops, base_eta, traffic_api):
    remaining_distance = calculate_distance(current_position, destination, via=remaining_stops)
    current_speed = traffic_api.get_current_speed(current_position, destination)
    historical_speed = get_historical_speed(current_position, destination, datetime.now())
    traffic_factor = historical_speed / current_speed
    remaining_time = (remaining_distance / historical_speed) * traffic_factor * 60
    return remaining_time

Comparison of approaches: Last Mile vs. Long Haul

Customer notifications

ETA is useless without integration with a communication layer:

Notification workflow:

1. After dispatch: "Your order is on its way, expected time: 14:30–15:00".
2. 60 minutes before: "Courier will arrive in ~55 minutes".
3. 15 minutes before: "Courier is nearby, will arrive in ~12 minutes".
4. If delay > 20% from ETA: automatic notification with new time and reason.

Channels: SMS (Twilio/SMS.ru), Push notifications, Email, WhatsApp Business API.

System metrics:

• ETA Accuracy: % of deliveries within ±15 min of ETA.
• ETA MAPE: average forecast error in percent.
• Proactive notification rate: % of delays that the customer was informed about before occurrence.
• CSAT correlation: correlation of ETA accuracy with delivery rating.

What's included in ETA system development

• Data audit: assess quality and completeness of historical data, configure pipelines.
• Feature engineering: develop a feature set tailored to your specifics (cargo type, region, seasonality).
• Modeling: LightGBM / LSTM / Quantile Regression, validation via cross-validation.
• Real-time update: integration with GPS tracker and traffic API.
• Notifications: configure triggers and channels (SMS, Push, Email).
• Metrics dashboard: panel for monitoring accuracy and proactivity.
• Support: documentation, training your team, 3-month warranty.

Timelines: basic ETA model with static forecast — 3–4 weeks. Real-time update + customer notifications + metrics — 10–12 weeks.

Order ETA system development — we will assess your project in 2 days. Get a consultation from our AI engineer. Contact us to discuss 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.