Time Series Model Training: Prophet, NeuralProphet, TimesFM

5+ years in AI/ML | 30+ forecasting projects | 500+ trained models. Backed by proven experience, we guarantee robust forecasting.

We develop and tune time series forecasting models — Prophet, NeuralProphet, and TimesFM — to help plan sales, purchases, and capacity utilization. Recently, one of our clients — a retail chain with 500 stores — faced unstable weekly sales forecasts: Seasonal Naive gave SMAPE of 15%, but the business required 10% accuracy. We chose Prophet From Wikipedia, the free encyclopedia, tuned changepoint_prior_scale=0.05, seasonality_mode='multiplicative', and achieved SMAPE of 8% — halving the error. This improvement saved the client $200,000 annually in inventory costs. This is a real example of how the right model and tuning solve the problem.

We have been in AI/ML for over 5 years and completed 30+ forecasting projects. Our approach includes strict cross-validation and comparison with multiple baselines. We publish model cards and ensure experiment reproducibility. Training projects typically start from $5,000 for a single model with cross-validation, up to $20,000 for a full production pipeline including monitoring and retraining. The budget is determined after analyzing your data — cost depends on data complexity and target accuracy.

What criteria should you use to select a model?

Choosing the right model depends on three factors: data volume, need for interpretability, and presence of external events (holidays, promotions). If you have little data (less than 2 years) and need transparency — pick Prophet. If there are nonlinear lags and external regressors — NeuralProphet. If you have abundant data (thousands of series) and need maximum accuracy without explanation — TimesFM. Contact us for a tailored recommendation based on your specific case.

What's included in end-to-end model training?

We prepare data, tune hyperparameters, perform cross-validation, build a production pipeline (Airflow + MLflow), document the model, and deliver an API endpoint. You get a trained model, documentation, an operational manual, and one month of support. Additionally, we set up data drift monitoring and automatic model retraining.

Prophet: Decomposition Model

Meta Prophet is an additive model:

y(t) = trend(t) + seasonality(t) + holidays(t) + ε(t)

Trend is piecewise linear or logistic growth. Changepoints — automatic detection of trend change points via L1 regularization.

Seasonality is described by Fourier series: yearly with N=10 (default), weekly with N=3, custom for any period.

Training and tuning:

from prophet import Prophet
import pandas as pd

m = Prophet(
    changepoint_prior_scale=0.05,   # trend flexibility
    seasonality_prior_scale=10.0,   # seasonality flexibility
    holidays_prior_scale=10.0,
    seasonality_mode='multiplicative'  # for data with growth
)
m.add_country_holidays(country_name='RU')
m.add_seasonality(name='monthly', period=30.5, fourier_order=5)
m.fit(df)  # df with columns ds, y

Key tuning parameters: changepoint_prior_scale (0.001–0.5) controls overfitting to trend, seasonality_mode ('additive' for stationary, 'multiplicative' for growing), fourier_order (higher = more flexible seasonality).

Prophet cross-validation:

from prophet.diagnostics import cross_validation, performance_metrics
df_cv = cross_validation(m, initial='730 days', period='180 days', horizon='365 days')
df_p = performance_metrics(df_cv)

NeuralProphet: Prophet with Neural Components

NeuralProphet extends Prophet with autoregression (AR-Net) and nonlinear lagged regressors. Training via PyTorch is significantly faster than MCMC-Prophet.

from neuralprophet import NeuralProphet

m = NeuralProphet(
    n_forecasts=7,        # forecast horizon
    n_lags=14,            # number of lags for AR
    seasonality_mode='auto',
    learning_rate=0.01
)
m = m.add_country_holidays('RU')
metrics = m.fit(df, freq='D', validation_df=df_val)

NeuralProphet is effective when nonlinear lag dependencies, multiple forecast steps, and external regressors with lag are present.

TimesFM: Foundation Model from Google

TimesFM is a pretrained foundation model for zero-shot forecasting. It requires no training on your data; first forecast in minutes.

import timesfm

tfm = timesfm.TimesFm(
    context_len=512,
    horizon_len=128,
    input_patch_len=32,
    output_patch_len=128,
    num_layers=20,
    model_dims=1280,
    backend='gpu'
)
tfm.load_from_checkpoint(repo_id="google/timesfm-1.0-200m")

# Zero-shot inference
forecast_input = [np.array(historical_data)]
frequency_input = [0]  # 0=high freq, 1=low freq
point_forecast, experimental_quantile_forecast = tfm.forecast(
    forecast_input, freq=frequency_input
)

Advantages: zero-shot, speed, strong results on most tasks. Limitations: does not account for known future covariates without fine-tuning, limited interpretability, requires significant context (>500 points).

Comparative Model Selection

Typical Timelines and Work Volume

Training and Validation Practice

Pipeline for training (example with Prophet):

1. Data preparation: resample to required frequency, handle missing values (interpolation / forward fill).
2. Analysis: ACF/PACF, decompose, holiday analysis.
3. Baseline: Seasonal Naive (forecast = value one year ago).
4. Prophet fit with default parameters, cross-validation.
5. Hyperparameter search: Optuna / grid search over 4–6 parameters (e.g., changepoint_prior_scale, seasonality_prior_scale, fourier_order).
6. Ensemble with competitor (ETS or NeuralProphet).
7. Productionization: Airflow DAG, MLflow tracking, API endpoint.

Metrics for comparison: SMAPE, MASE (normalized on Seasonal Naive), Winkler Score for interval forecasts.

Timelines: tuning and training Prophet/NeuralProphet for a single series with cross-validation — 1–2 weeks. Production pipeline with monitoring and auto-retraining — 4–6 weeks. Savings from accurate forecasting can cover the project cost within the first months. Get an individual offer — contact us for a consultation on your task. We will evaluate the project and offer an optimal solution.

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.