High Performance Time Series Forecasting in Python
Forecast what the business actually plans against. In five days you will build production forecasting systems in Python — ARIMA, Prophet, gradient-boosted and hierarchical models — on real demand data, and defend a forecast to the finance and supply-chain leaders who live with its errors.
The executive challenge
The problem this program solves
Most forecasting courses stop at a fitted model; this one starts where forecasting matters — at the meeting where a forecast error becomes a stock-out, a cash squeeze or a missed quarter. You will build and backtest complete forecasting systems in Python on African demand data, learn the backtesting discipline that exposes overfit forecasts, and finish defending your capstone forecast to a mock S&OP panel. Taught by faculty led by Dr. Kennedy Mwai (Roche), who build forecasting pipelines for regulated, high-stakes data.
Context
Why this matters now
Capability
What you will be able to do
- Decompose and diagnose real time series — trend, seasonality, holiday and structural breaks — before choosing a model
- Build classical forecasts with exponential smoothing and ARIMA/SARIMA in Python with proper statistical diagnostics
- Apply Meta’s Prophet and modern AutoML forecasting libraries to multi-season business series
- Engineer lag and rolling features for gradient-boosted demand models — the current workhorse of retail forecasting
- Backtest forecasts correctly with rolling-origin evaluation and loss metrics matched to the business decision
- Produce hierarchical and reconciled forecasts, and communicate uncertainty as ranges, not single numbers
The IPAT Course Standard
The five-day journey
Each day is a deliberate progression in managerial capability — from strategic context, to analytical frameworks, to applied practice, to executive judgement under pressure, to a capstone that produces something your organisation keeps.
IPAT Signature Case™
The Stock-Out That Wasn’t in the Model: FMCG Demand at Distribution Scale
A leading East African FMCG distributor forecast national demand accurately at aggregate level — and still ran out of stock in the regions that mattered. The aggregate series hid a bimodal pattern: urban SKUs with stable weekly seasonality, upcountry SKUs with intermittent, promotion-driven demand. One national model was quietly right about the average and wrong about every distribution centre that counted.
Your decision: Invest in a hierarchical per-segment forecasting system requiring new data engineering, extend the single national model with segment-level adjustments, or accept the aggregate accuracy and absorb stock-outs with buffer stock — trading engineering cost, service level and working capital.
Real-World Case Study
The M5 Competition: What Thousands of Teams Taught the Forecasting World
The M5 forecasting competition (Makridakis et al., 2020–2022) asked thousands of teams to forecast Walmart sales across 42 US states, 3,049 products and hierarchical levels. The results reshaped practice: gradient-boosted tree methods dominated, hierarchical reconciliation mattered as much as model choice, and the best teams spent their effort on features and validation design rather than exotic architectures.
The lesson: Forecasting performance is won in validation design and feature engineering — the exact disciplines this course drills — not in chasing the newest model family.
IPAT Data Lab™
A distributor’s finance team plans cash against a forecast that is accurate on paper and useless at the depot level — and nobody can say which model deserves the budget.
Dataset: A realistic multi-region, multi-SKU demand dataset with weekly seasonality, promotions, holiday effects, intermittent upcountry series and known structural breaks.
Your task: Build the full forecasting system in Python — diagnosis, benchmarks, ETS/SARIMA, Prophet and gradient boosting under rolling-origin backtesting — then reconcile the hierarchy and price the service-level impact.
A defended forecasting system with a backtested leaderboard and reconciled portfolio forecast.
Data → Analysis → Decision, not Data → Code → Certificate
IPAT DIOS™ Framework · Data Intelligence for Official Statistics & Organisations
From raw data to strategic intelligence in five phases
Software-led programmes at IPAT do not teach syntax — they teach statistical data architecture, the discipline that turns high performance time series data into a strategic asset. The programme runs on the five-phase IPAT DIOS™ arc, adapted end-to-end to Python & pandas.
Assess
Diagnose the current state of high performance time series data processing: a baseline of Python & pandas proficiency, data silos, manual bottlenecks and reproducibility gaps in the workflows your institution runs today.
Architect
Design standardized, reproducible data structures: consistent naming conventions, master schemas and version-controlled projects in Git, so every high performance time series analysis stands on the same interoperable foundation.
Automate
Replace manual, error-prone steps with automated pipelines: ingest, clean and transform high performance time series data with pandas, NumPy and SciPy, and automate with reproducible Jupyter and Quarto pipelines the routine reporting that consumes weeks of analyst time.
Analyze
Create decision intelligence: cross-tabulation with proper variance estimation, time-series visualization, geospatial mapping and predictive modelling — applied to realistic, anonymized high performance time series data.
Accelerate
Build for what is next: AI-ready datasets, machine learning for imputation and estimation, and interactive Streamlit and Plotly dashboards that put high performance time series intelligence directly in decision-makers’ hands.
Day 5 · The Capstone
What will you change when you return to work?
Build a twelve-month analytical and AI roadmap for your organisation: the value case, the data foundations, the governance, and the first 90 days.
Methodology
Frameworks you will master
- • IPAT DIOS™ Decision-Integrity Standard
- • statsforecast / Nixtla toolchain
- • Prophet decomposable forecasting
- • Rolling-origin backtesting
- • Hierarchical forecast reconciliation
Career Outcomes
Where this program takes you
Salary bands reflect 2025–26 market benchmarking across African markets plus global mobility ranges.
Cohorts
Upcoming cohorts
New dates are being scheduled
Register with your preferred start date and our admissions team will match you to the next available cohort — or contact us to arrange an in-house delivery for your organisation.
Register your interestExplore further
Related programmes
Learners of High Performance Time Series Forecasting in Python frequently pair it with these offerings from the same divisions.