All projects
data
A forecast with honest error bars
Predict a real series, and be clear about how wrong you might be.
Why it matters
Forecasting is where naive evaluation does the most damage, because a random train-test split leaks the future. Getting the backtest right is the whole skill.
What you build
- A naive baseline to beat
- Backtesting that respects time order
- Prediction intervals, not just points
- An error metric appropriate to the series
If you want more
- Add multiple seasonalities
- Add holiday effects
Build it
Surprisingly hard to beat. That is the point of doing it first.
- Choose and justify the error metric
Material for the whole build
Shows you understand temporal leakage, which is the mistake that makes most forecasting portfolios worthless.
Get a plan built around projects like this