Time Series Forecasting Implementation

Time series forecasting with Prophet, N-BEATS, and LightGBM requires careful walk-forward validation to avoid data leakage. We regularly encounter projects where data—sales, IoT sensors, or stock quotes—contains temporal dependencies easily broken by improper handling. An incorrect split or ignoring seasonality leads to data leakage and falsely optimistic backtest results. For instance, in one spare parts demand forecasting project, seasonal naive gave MAPE of 40%, Prophet 28%, but after walk-forward validation Prophet was 10% worse on last three months. Over years of real-world projects, we have built over 20 forecasting systems and developed a robust methodology. For each task, we select appropriate stack—from classic SARIMA to modern Transformers—considering budget and interpretability. We analyze your data and propose a roadmap in 2–3 days.

In one retail project with 5000 SKUs, we implemented LightGBM with walk-forward validation and quantile forecasts, reducing MAPE from 35% to 18% and saving the client €1M annually in inventory costs. Our pricing starts at €5k for baseline models and goes up to €30k for full MLOps pipeline.

Time Series Forecasting with Prophet, N-BEATS, and Walk-Forward Validation

Before choosing a method, we analyze series properties:

• Stationarity: ADF test. Non-stationary series require differencing or specialized methods.
• Seasonality: ACF/PACF analysis. Single (weekly) or multiple (weekly + yearly) influences model selection.
• Intermittency: ADI > 1.32 calls for special methods (Croston, IMAPA).
• Nonlinearity: Teräsvirta test or BDS test. Linear models (ARIMA) inadequate under strong nonlinearity.

There is no universal answer—we compare candidates on historical data. Typical options with trade-offs:

• Naive / Seasonal Naive — simplest baseline to verify complex methods outperform.
• ETS (Exponential Smoothing) — works well on series with single seasonality.
• SARIMA — classic with confidence intervals, but slow for large observations.
• Prophet — convenient for business data with holidays, interpretable, but loses to neural networks on complex patterns.
• LightGBM with lags — high accuracy with many external factors, but requires engineering work on features.
• N-BEATS / N-HiTS — SOTA on M4/M5 competitions, works without external features.
• Temporal Fusion Transformer — leader for ensembles of multiple series, but demands GPU.
• TimesGPT / TimesFM — foundation models for zero-shot forecasting, accelerate start.

For example, LightGBM trains 5x faster than N-BEATS on medium datasets (10k rows), yet N-BEATS achieves 1.5x lower MAPE on complex seasonal patterns. We choose based on your data.

How to Avoid Data Leakage in Time Series Forecasting?

Standard train/test split violates temporal ordering.

Walk-Forward Validation:

    |---Train---| Test |
    |----Train----|   Test   |
    |-----Train-----|     Test     |
Average metrics across all windows

Test window size = forecast horizon. Shift step = horizon / 2 or horizon (no overlap).

Data leakage sources:

• Using future data in scaling (fit scaler on entire dataset)
• Target encoding with future values
• External features with future information (known future covariates vs. past covariates)

Without it, any test metric is optimistic. We guarantee all models undergo temporal split without overlap. Using statsforecast library with automatic window selection, we obtain realistic quality estimates and avoid overpaying for false expectations.

Walk-forward with overlap gives most stable metrics and matches production load.

Feature Engineering and Probabilistic Forecasting

Temporal features:

df['hour'] = df.index.hour
df['day_of_week'] = df.index.dayofweek
df['week_of_year'] = df.index.isocalendar().week
df['month'] = df.index.month
df['is_weekend'] = df['day_of_week'].isin([5, 6]).astype(int)
df['sin_hour'] = np.sin(2 * np.pi * df['hour'] / 24)
df['cos_hour'] = np.cos(2 * np.pi * df['hour'] / 24)

Lag features: t-1, t-7, t-14, t-28 for daily data; t-1, t-24, t-168 for hourly.

Rolling statistics: mean, std, min, max over 7/28/90 days. Differences: (t-1) - (t-7) to capture trend.

A point forecast without uncertainty is insufficient. Quantile forecasts give intervals. Example: LightGBM with objective='quantile', alpha=0.1/0.5/0.9. Conformal prediction provides theoretically grounded intervals. N-HiTS in neuralforecast library supports quantiles natively.

Using quantile forecasts, a retail client reduced inventory costs by 20% compared to point forecasts—saving €200k annually.

What Is a Quantile Forecast?

A quantile forecast provides an uncertainty interval instead of a point value. For example, forecast of 100 units with percentiles [70, 130] allows risk-aware decisions. Implemented via quantile regression (LightGBM), conformal prediction, or Monte Carlo Dropout.

Production Pipeline

from statsforecast import StatsForecast
from statsforecast.models import AutoARIMA, AutoETS, AutoTheta

models = [AutoARIMA(season_length=7), AutoETS(season_length=7), AutoTheta()]
sf = StatsForecast(models=models, freq='D', n_jobs=-1)
sf.fit(train_df)
forecasts = sf.predict(h=28, level=[80, 95])

MLflow tracks each experiment: data version, hyperparameters, metrics, model artifact. Airflow DAG schedules daily retraining and publishes forecasts to Data Warehouse. Evidently monitors data drift and prediction drift.

Implementation Plan and Deliverables

1. Analyze raw data, identify patterns (stationarity, seasonality, intermittency).
2. Select baseline and 3–5 candidates (from simple to complex).
3. Walk-forward validation of each model, compare by MAPE, RMSE, MASE.
4. Develop production pipeline: MLflow, Airflow, Evidently.
5. Integrate forecasts into Data Warehouse, set up alerts.

Timeline: 2–3 weeks for baseline (€5k–€10k), 8–12 weeks for full system (€15k–€30k). Contact us for a project assessment—we will evaluate your time series and prepare a roadmap in 2–3 days.

Our team's experience includes over 20 completed projects, average MAPE reduction of 15–30% after tuning. Reach out to get a detailed commercial proposal.

Deliverables

• Data analysis report with stationarity, seasonality, and nonlinearity tests.
• Selection and tuning of 3–5 candidate models, compared via walk-forward validation.
• Production-ready pipeline: MLflow tracking, Airflow orchestration, Evidently monitoring.
• Documentation: model cards, data schemas, deployment guide.
• API or database integration for forecasts.
• Training session for your team (2 hours).
• 3 months of support and model retuning.

In time series forecasting, walk-forward validation is essential for honest evaluation. For time series forecasting projects, we recommend starting with a baseline like Prophet and scaling up to LightGBM or N-BEATS for better accuracy. Our clients typically see ROI within 6 months, with savings ranging from €100k to €1M.

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.