FORMATION IA
Prévision de Séries Temporelles pour les Équipes Opérationnelles
Construisez des pipelines de prévision fiables, des tableurs aux modèles ML prêts pour la production.
Ce qu'elle couvre
Ce programme de niveau praticien enseigne aux analystes des opérations, de la supply chain et de la finance comment concevoir et déployer des systèmes de prévision de séries temporelles. Les participants progressent des méthodes statistiques classiques (ARIMA, lissage exponentiel) vers des approches ML modernes (Prophet, LightGBM, TimeGPT, Nixtla) et apprennent à évaluer les modèles de manière rigoureuse à l'aide de frameworks de backtesting. Le cours comprend des travaux pratiques pour construire des pipelines de prévision de bout en bout et intégrer les résultats dans les flux de décision opérationnelle.
À l'issue, vous saurez
- Select the appropriate forecasting method (statistical vs. ML vs. foundation model) for a given operational problem and dataset
- Implement a backtesting harness to evaluate and compare forecast accuracy using MAPE, RMSE, and coverage metrics
- Build a Prophet or Nixtla StatsForecast pipeline with custom seasonalities and external regressors from raw operational data
- Deploy a forecasting model as a scheduled pipeline with automated retraining triggers and drift monitoring
- Communicate forecast uncertainty to non-technical stakeholders using confidence intervals and scenario ranges
Sujets abordés
- Classical forecasting: ARIMA, Exponential Smoothing, Holt-Winters
- ML-based forecasting: LightGBM, XGBoost with lag features
- Foundation models: Prophet, TimeGPT, Nixtla StatsForecast
- Backtesting and cross-validation for time-series evaluation
- Feature engineering: seasonality, holidays, external regressors
- Forecast uncertainty and confidence intervals
- Production pipeline design: scheduling, monitoring, retraining
- Integrating forecasts into operational dashboards and planning tools
Modalité
Delivered as a blended programme over 3–5 weeks: two live instructor-led sessions per week (90 minutes each) combined with async lab work between sessions. Labs use Jupyter notebooks with real-world datasets (retail demand, energy consumption, financial revenue). Participants bring one internal dataset to apply learnings directly. Hands-on ratio is approximately 60% labs, 40% instruction. Remote delivery via video conferencing; in-person cohort delivery available on request for groups of 10+.
Ce qui fait que ça marche
- Anchoring every model choice to a business metric (e.g., inventory holding cost, stockout rate) rather than pure statistical accuracy
- Establishing a baseline naive forecast at the outset so improvement is always measurable and communicable
- Involving end-users (planners, buyers, finance) in reviewing forecast outputs during the programme, not just at the end
- Automating retraining and monitoring from day one so the pipeline remains operational without manual intervention
Erreurs fréquentes
- Jumping straight to complex ML models without establishing a statistical baseline, making it impossible to measure real improvement
- Using random train/test splits instead of time-ordered backtesting, leading to optimistic and invalid accuracy scores
- Ignoring forecast uncertainty and presenting point estimates to planners who then make binary go/no-go decisions on them
- Building a one-shot forecast model with no retraining schedule, causing silent degradation as patterns shift
Quand NE PAS suivre cette formation
This training is not the right fit for teams that do not yet have clean, consistent historical data at the required granularity — if your organisation cannot export 18+ months of reliable transaction or operational records, the value of advanced forecasting methods will be severely limited and a data quality initiative should come first.
Fournisseurs à considérer
Sources
Cette formation fait partie d'un catalogue Data & IA construit pour les leaders sérieux sur l'exécution. Lancez le diagnostic gratuit pour voir quelles formations sont prioritaires pour votre équipe.