Change one thing.
Add a signal or change the forecasting method while keeping the other rules fixed within that study.
Research contribution · backtesting and evaluation
Testing macro signals,
from forecasts to portfolios.
We built a framework to test whether new macro information and forecasting methods improve investment decisions. Within each experiment, we use shared dates and portfolio rules so we can see what each change adds.
Learn from earlier data.
Train on completed outcomes, fit preprocessing inside each training window, and trade after the decision.
Follow the whole result.
Measure prediction errors, then examine portfolio returns, losses, trading costs and consistency across periods.
Economic histories contain later revisions and some availability dates are assumed. These development experiments are exploratory, not a complete real-time reconstruction or a final test.
The research question
What improves the decision—and what only changes it?
A stronger research result requires more than a higher backtest return. We compare forecasts with a simple past-average prediction and portfolios with common reference strategies.
Completed · exploratory
Does adding a signal help?
Test unemployment claims, credit spreads and the yield curve separately against the original wider-input model.
Explore added signals → Completed · matched comparisonDoes the forecasting method matter?
Compare the historical mean, OLS, Ridge and Elastic Net, then test lags and interactions as separate Ridge extensions.
Compare forecasting methods → Planned extensionDoes the forecast horizon matter?
Current experiments predict one-month returns. Three- and six-month return targets are not yet implemented or evaluated.
Longer evaluation windows are not longer forecast horizons.See the horizon-testing plan →What we learned so far
The framework reveals tradeoffs, not just winners.
The studies use different information sets and samples. Read each result within its own experiment.
Extra information needs to earn its place.
These initial signal runs use the original model, whose forecasts were unstable. Added signals have not yet been rerun on the corrected reference model.
Judge predictions and portfolios separately.
Each forecasting method is scored on prediction errors and on the portfolio it produces. A high return alone does not establish forecast skill, and a better forecast does not automatically improve risk or trading costs.
Inspect both levels of evidence →Reliability checks can change the conclusion.
Read the reliability case study →What remains
A clear boundary between findings and next steps.
The framework is the contribution; the current findings show where the evidence is strong, weak or unfinished.
Keep the initial runs visible as exploratory evidence. A new signal comparison needs its own recorded rules and matched dates.
Three- and six-month tests need explicit target, training, overlap and portfolio rules before they can produce valid comparisons.
Collect historical vintages and verified release dates where possible. Timing checks do not remove revised-data bias.
The chosen reference model’s rules are locked. Confirm whether 2023 onward remains unexamined before a separately authorized evaluation.