Skills

Build a Tested Time-Series Forecast

4 weeks · 0 milestones

Turn real market or macro data into a clean return series and beat — or honestly fail to beat — a naive baseline on data the model never saw.

Part of: Business & Management · Proves for: Career

Milestone map

3 milestones

Build a clean return series

1 week

Download real market or macro data, convert prices to returns, and check for gaps, outliers and stationarity. Every later result depends on this data being right.

Proof required

Submit the notebook that loads the data, cleans it and plots returns, with a note on gaps, outliers and a stationarity check.

What gets checked

Real data with source and dates
Returns, not raw prices, modelled
Cleaning choices written down

Common mistakes

Modelling prices instead of returns
Silently dropping bad rows
No data source

Powstik Guide

A clean, documented return series.

Steps

Pick a free data source.
Load and align dates.
Compute log or simple returns.
Check gaps and outliers.
Run a stationarity test and note the result.

Template

Source:
Dates:
Frequency:
Gaps handled:
Outliers:
Stationarity test result:

What gets sent back

No source.
Price-level modelling.
Undocumented cleaning.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Check the data source and date range
Ask why returns and not prices

You'll sign in first, then come straight back here.

Fit and test out of sample

2 weeks

Fit a baseline and one better model (for example ARIMA). Test only on later data the model never saw, using a rolling time-based split.

Proof required

Submit out-of-sample error for the baseline and your model on a time-based split, with the code.

What gets checked

Beats or honestly fails to beat a naive baseline
Test data strictly after training data
Error metric stated

Common mistakes

Shuffled cross-validation on time series
No baseline
Reporting in-sample fit

Powstik Guide

An honest out-of-sample comparison.

Steps

Make a naive baseline.
Fit your model on the early part.
Roll forward and predict the next period.
Compare errors.
Say plainly if the baseline wins.

Template

Baseline error:
Model error:
Split method:
Winner:

What gets sent back

Shuffled split.
No baseline.
In-sample results only.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Check no future data leaks into training
Ask how the model compares to 'tomorrow = today'

You'll sign in first, then come straight back here.

Defend the model live

1 week

Walk a reviewer through your data, model and results. They pick a new time period or asset and you rerun the test live.

Proof required

Submit the rerun on the reviewer's chosen period and the review log.

What gets checked

Rerun works on a period you did not choose
Limits of the model named
Reviewer challenges answered

Common mistakes

Model only works on the tuned period
Reviewer only watched
Overstating accuracy

Powstik Guide

A model that survives a new period.

Steps

Book a session.
Let the reviewer choose a period or asset.
Rerun live.
Discuss where it fails.
Log the session.

Template

Review session log
Date:            Duration:
Reviewer name:
Reviewer role and experience:

Challenge 1 (time-series model on a new period):
  What they asked:
  My answer:
Challenge 2:
  What they asked:
  My answer:

Change I made after the session:

What gets sent back

No live rerun.
No new period.
No log.

Resources

Foundationstart here

Depthgo deeper

Masteryfor the dedicated

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Pick a new period or asset
Ask where the model breaks

You'll sign in first, then come straight back here.

Part of

Quant Modeling →

Related outcomes

Milestone map

3 milestones

Build a clean return series

1 week

Download real market or macro data, convert prices to returns, and check for gaps, outliers and stationarity. Every later result depends on this data being right.

Proof required

Submit the notebook that loads the data, cleans it and plots returns, with a note on gaps, outliers and a stationarity check.

What gets checked

Real data with source and dates
Returns, not raw prices, modelled
Cleaning choices written down

Common mistakes

Modelling prices instead of returns
Silently dropping bad rows
No data source

Powstik Guide

A clean, documented return series.

Steps

Pick a free data source.
Load and align dates.
Compute log or simple returns.
Check gaps and outliers.
Run a stationarity test and note the result.

Template

Source:
Dates:
Frequency:
Gaps handled:
Outliers:
Stationarity test result:

What gets sent back

No source.
Price-level modelling.
Undocumented cleaning.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Check the data source and date range
Ask why returns and not prices

You'll sign in first, then come straight back here.

Fit and test out of sample

2 weeks

Fit a baseline and one better model (for example ARIMA). Test only on later data the model never saw, using a rolling time-based split.

Proof required

Submit out-of-sample error for the baseline and your model on a time-based split, with the code.

What gets checked

Beats or honestly fails to beat a naive baseline
Test data strictly after training data
Error metric stated

Common mistakes

Shuffled cross-validation on time series
No baseline
Reporting in-sample fit

Powstik Guide

An honest out-of-sample comparison.

Steps

Make a naive baseline.
Fit your model on the early part.
Roll forward and predict the next period.
Compare errors.
Say plainly if the baseline wins.

Template

Baseline error:
Model error:
Split method:
Winner:

What gets sent back

Shuffled split.
No baseline.
In-sample results only.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Check no future data leaks into training
Ask how the model compares to 'tomorrow = today'

You'll sign in first, then come straight back here.

Defend the model live

1 week

Walk a reviewer through your data, model and results. They pick a new time period or asset and you rerun the test live.

Proof required

Submit the rerun on the reviewer's chosen period and the review log.

What gets checked

Rerun works on a period you did not choose
Limits of the model named
Reviewer challenges answered

Common mistakes

Model only works on the tuned period
Reviewer only watched
Overstating accuracy

Powstik Guide

A model that survives a new period.

Steps

Book a session.
Let the reviewer choose a period or asset.
Rerun live.
Discuss where it fails.
Log the session.

Template

Review session log
Date:            Duration:
Reviewer name:
Reviewer role and experience:

Challenge 1 (time-series model on a new period):
  What they asked:
  My answer:
Challenge 2:
  What they asked:
  My answer:

Change I made after the session:

What gets sent back

No live rerun.
No new period.
No log.

Resources

Foundationstart here

Depthgo deeper

Masteryfor the dedicated

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Pick a new period or asset
Ask where the model breaks

You'll sign in first, then come straight back here.