Machine learning & data science
Forecast time series
Seasonality, ARIMA/Prophet/gradient boosting, backtesting for demand/finance use cases.
~12 focused hours·intermediate
Tools: Prophet, statsmodels (ARIMA/ETS), scikit-learn/XGBoost, pandas
Market relevance — share of job ads asking for this
Prerequisite capability — not asked for directly, but needed for others.
What employers mean
You should be able to…
- Decompose a time series into trend, seasonality and residual components
- Build a classical statistical forecast (ARIMA/ETS/Prophet) with confidence intervals
- Build a machine-learning forecast (feature-engineered gradient boosting) and compare it to the classical baseline
- Backtest a forecasting model correctly with a rolling/expanding time-based split, never a random split
- Handle holidays, promotions and other known events that shift demand
- Choose the right evaluation metric (MAPE, RMSE, WAPE) for a business forecasting problem
- Communicate forecast uncertainty (intervals, not just a point estimate) to a business stakeholder
Needs first: Train and evaluate classical ML models
Learn — free, link-checked
The few resources that matter
Read · beginner · 30 min · facebook.github.io
Quick Start
Official Prophet quickstart -- fit a seasonal forecast with holidays and uncertainty intervals in under an hour, a common demand-forecasting baseline. — Meta (Facebook)
Read · intermediate · 60 min · tensorflow.org
Time series forecasting
Shows windowing, feature engineering and comparing a naive baseline against RNN/CNN forecasters -- the backtesting mindset interviewers want. — TensorFlow
Course · intermediate · 180 min · kaggle.com
Time Series
Hands-on trend/seasonality/cycle decomposition and hybrid model notebooks with an immediate competition-style evaluation. — Kaggle Learn
Read · intermediate · 300 min · otexts.com
Forecasting: Principles and Practice (3rd ed)
The standard free forecasting textbook -- seasonality decomposition, ARIMA, ETS and backtesting explained by the authors of the R forecast/fable packages. — Rob J Hyndman & George Athanasopoulos
Practice
Festival-season demand forecast for an Indian retail/NBFC dataset
Using a public retail sales dataset (or an NBFC loan-disbursal proxy series) with visible seasonality around festivals like Diwali, build a Prophet baseline with Indian holiday effects included, then a feature-engineered XGBoost forecaster, and backtest both with a rolling time-based split. Report WAPE/MAPE and visualize forecast vs actuals with confidence intervals.
Done when
- Prophet model includes an Indian holiday/festival calendar as regressors, not just default seasonality
- XGBoost (or similar) forecaster built with lag/rolling features and compared against the Prophet baseline
- Backtesting uses a rolling or expanding time-based split across multiple folds, with per-fold WAPE/MAPE reported
- README includes a plot of forecast vs actuals with confidence intervals and states which model won and why
Prove it
Evidence a recruiter can check
- Public GitHub repo with backtesting code showing the time-based split logic explicitly
- A results table comparing WAPE/MAPE across models and backtest folds
- Forecast-vs-actual plots with confidence intervals checked into the README
- Can explain live why a random train/test split would have been invalid here
Interview
Questions you'll get asked
- Walk me through how you'd forecast next month's demand for a product with strong seasonality
- Why must you backtest a forecasting model with a time-based split instead of random cross-validation?
- ARIMA vs Prophet vs a gradient-boosted regressor for forecasting -- when would you pick each?
- How do you incorporate known future events (festivals, promotions) into a forecast?
- What's the difference between MAPE and WAPE, and when does MAPE break down?
- How would you communicate forecast uncertainty to a non-technical stakeholder?
- How do you detect and handle a structural break (e.g. COVID-like shock) in historical demand data?