Backtesting without time travel
A faithful backtest is a replay of past information states, not merely a sorted split.
Regional demand shifts through promotions, holidays, and drift.
This chapter moves from mirror the release cadence to keep the distribution of outcomes. We will make the idea concrete in a small code example before placing it back inside a production system with available before staffing decisions.
Report error by origin and horizon with uncertainty; a mean alone hides whether one regime accounts for the apparent gain.
Verify the behavior illustrated in rolling_origins.py, then measure decision utility, slice reliability, and system latency.
Training uses realized future weather and random folds, producing a forecast impossible to reproduce at origin time. An unsafe release must degrade to a named baseline or the last known-good version.
Build the idea one layer at a time.
One train/test split tells one historical story. Rolling-origin backtesting asks the production question repeatedly: what would we have predicted from each past decision point using only knowledge then available?
At every origin, training data must end before the validation horizon begins. Expanding windows retain all history; sliding windows limit training to a recent regime. A gap may also be required when labels mature late or feature computation crosses the split boundary.
Preprocessing and tuning live inside each backtest fold. Fitting a scaler on all dates, choosing hyperparameters from the final window, or materializing rolling features before splitting can leak the future even when row indices appear chronological. Aggregate fold results only after retaining horizon and slice detail.
- rolling backtests
- direct vs recursive
- quantiles
- coverage
Follow the reasoning before touching the code.
- 01
Mirror the release cadence
Choose origins, training-window length, gap, and forecast horizon to match how often production retrains and how far ahead it predicts.
- 02
Rebuild state per origin
Refit transforms, feature statistics, and model parameters using only rows whose values and labels were mature at that origin.
- 03
Keep the distribution of outcomes
Report error by origin and horizon with uncertainty; a mean alone hides whether one regime accounts for the apparent gain.
def rolling_origins(size, first_origin, horizon, step):
for origin in range(first_origin, size - horizon + 1, step):
train = list(range(0, origin))
valid = list(range(origin, origin + horizon))
yield origin, train, valid
for origin, train, valid in rolling_origins(12, 6, 2, 2):
assert max(train) < min(valid)
print('origin', origin, 'train', train, 'valid', valid)Generate expanding-window train and validation indices.
What to notice
- Each validation window sits strictly after its training rows, reproducing the direction in which production forecasts move.
- Later folds have more training history; use a sliding window instead if the deployed model deliberately forgets old regimes.
YOUR TURN Pause and predict before revealing the answer.
Add a one-period label-maturity gap between training and validation. What should the last training index be for origin 6?
Need a hint?
An index immediately before origin is not yet eligible when the gap is one.
Check your reasoning
The last training index should be 4, leaving index 5 excluded as the maturity gap before validation starts at 6.
Lessons 1, 2 in this course.
Connect the chapter to the course’s quantitative spine.
A faithful backtest is a replay of past information states, not merely a sorted split. The quantile forecast objective below is the shared course-level lens that helps quantify the surrounding system. Name each symbol, check its units or shape, and use only values that were available at the production cutoff.
Quantile forecast objective
| Symbol | Meaning / shape / unit |
|---|---|
τ | target quantile |
q̂_τ | predicted quantile |
y | observed value |
Open derivation and numerical substitution
Start from the production quantity being optimized, substitute the observed values with their declared units, then isolate the model-controlled term. Preserve shape annotations at each step so broadcasting or aggregation cannot silently change the result.
- Write the named inputs: τ, q̂_τ, y.
- Substitute one small, hand-checkable batch before vectorizing.
- Calculate an independent reference value and compare within a declared tolerance.
# equation → code contract
inputs = validate_shapes_and_units(batch)
value = compute_c07(inputs)
assert is_finite(value)Change one assumption and make the tradeoff visible.
Use the worked chapter as your prediction, then change one input in the lab and explain why the output moves. The lab runs predefined TypeScript and never executes learner code, so the focus stays on the concept rather than setup.
Rolling-window backtest
Change history length and inspect bias, error, and interval coverage.
Assumption: Seasonal demand with one simulated regime change at week 30.
Open nonvisual data table
| Item | Computed state | Interpretation |
|---|---|---|
| h+1 | 12.5 MAE | 87.8% |
| h+6 | 13.6 MAE | 86.7% |
| h+12 | 15.0 MAE | 85.4% |
| h+18 | 16.3 MAE | 84.0% |
| h+24 | 17.6 MAE | 82.7% |
Trace the complete operating path.
Carry the chapter’s central idea into the system: A faithful backtest is a replay of past information states, not merely a sorted split. The path below shows where that reasoning needs an artifact, a test, an owner, or a fallback.
- 01
Turn “Mirror the release cadence” into a written invariant and a failing test.
- 02
Promote rolling_origins.py from illustration to a versioned implementation with representative fixtures.
- 03
Use “Keep the distribution of outcomes” as the release checkpoint: Report error by origin and horizon with uncertainty; a mean alone hides whether one regime accounts for the apparent gain.
- 04
Serve or materialize behind explicit available before staffing decisions.
- 05
Join telemetry to mature outcomes and retain a rollback path.
Observability
Join service health, input quality, prediction distributions, slice behavior, and mature outcomes by exact version.
Cost
Measure storage, preprocessing, compute, queueing, and human review under a representative arrival pattern.
Failure modes
Training uses realized future weather and random folds, producing a forecast impossible to reproduce at origin time. Add a detector, owner, mitigation, and stop condition for this class of failure.
Alternatives
Compare a rule, a simpler statistical baseline, and a different system boundary before adding model complexity.
Explain the contract, not just the vocabulary.
Regional demand forecast
Build a 24-hour pipeline with naive, statistical, and boosted baselines plus interval forecasts.
- Forecast contract
- Rolling-origin evaluation
- Horizon and coverage report
- Drift/retrain/fallback policy
Read primary material with a purpose.
Forecasting: Principles and Practice
90 min · last checked 26 Aug 2026
↗RECOMMENDED · Official docsscikit-learn lagged forecasting example
25 min · last checked 26 Aug 2026
↗DEEP DIVE · INDEXFind a primary implementation or benchmark
30–90 min · verify task, dataset, and evaluation match
↗Return to the opening failure.
Monitor horizon-level error, bias, interval coverage, freshness, reconciliation, and a seasonal-naive fallback.
For this lesson, the release evidence is the worked code example, a hand-checked formal result, deterministic simulation output, a ≥80% checkpoint, the production rubric, and a named fallback. The course resolves when the system can produce rolling validation, statistical baselines, boosted forecasts, and uncertainty bands.