Forecast horizons · planned extension
Does the answer change
when we look further ahead?
That is a useful next research question. The current deterministic experiments predict one-month asset returns. Three- and six-month return targets have not yet been implemented or backtested.
This page records the proposed extension. There are no longer-horizon results or active horizon controls to interpret as completed tests.
Keep these three ideas separate
A forecast horizon is the period being predicted.
Several parts of the existing study mention months, but they refer to different things.
| Term | What it means here | Example |
|---|---|---|
| Forecast horizon | The future return interval the model predicts | Next month’s SPY return · currently evaluated |
| Input lookback / policy expectation | A feature summarizing earlier changes or expected policy rates | Three-month claims change; six-month policy-rate expectation |
| Evaluation window | The period over which repeated decisions are scored | A COVID episode or rolling 12-month portfolio window |
A controlled next experiment
What must be specified before longer horizons can be compared?
A horizon selector is not enough. Targets, timing and portfolio rules must agree.
Define the return targets
Specify exact entry and exit prices for one-, three- and six-month predictions, along with a common feasible sample.
Respect target completion
A six-month return cannot enter training until all six months have finished. Tuning folds need the same rule.
Specify the trading strategy
Decide whether portfolios hold for the full horizon or rebalance monthly, how overlapping holdings work, and how views and covariance use matching units.
Evaluate overlapping outcomes honestly
Longer targets can overlap. Error summaries and uncertainty checks must account for that dependence.
Record a separate specification
The locked corrected reference is a one-month model. A new horizon study needs its own documented rules before running or inspecting its results.