Frequency
Weekly, Friday-ended
Daily source series are aligned, then resampled with W-FRI using the last available observation. Incomplete future Fridays are excluded.
Production status
What this methodology is — and is not
USD Impact Score v2 is a descriptive weekly regime indicator. It is not a return forecast, trading signal, probability model or optimized portfolio rule. Positive values describe conditions that the framework associates with firmer dollar pressure; negative values describe softer-dollar conditions.
The model is intentionally simple: eight market levels, fixed signed weights and z-score standardization. That simplicity makes the arithmetic auditable, but it does not remove specification risk, correlation risk or regime dependence.
Exact calculation
The production formula
For each driver i and week t, the pipeline takes the last available observation in the Friday-ended week and standardizes the weekly level against the complete production sample available at run time T:
zi,t,T = clip((xi,t − μi,T) / si,T, −3.5, +3.5)
μ is the full-sample mean and s is the pandas sample standard deviation (the default ddof=1) for all complete Friday observations from the production start date through run T.
Scoret,T = 0.125 × (DXY − WTI − SPX + VIX − BTC − GOLD + UST2Y + UST10Y)
Every term in the second formula is the clipped z-score above. The eight absolute weights sum to 1.00.
Variables and weights
What enters the model
| Driver | Production series | Input | Weight | Framework rationale |
|---|---|---|---|---|
| DXY | Yahoo: DX-Y.NYB | Weekly level | +0.125 | Direct dollar-index pressure. |
| WTI | Yahoo: CL=F | Weekly level | −0.125 | Commodity/reflation channel; higher oil is assigned a softer-dollar sign. |
| S&P 500 | Yahoo: ^GSPC | Weekly level | −0.125 | Risk/liquidity channel; stronger risk assets are assigned a softer-dollar sign. |
| VIX | Yahoo: ^VIX | Weekly level | +0.125 | Stress channel; higher volatility is assigned a firmer-dollar sign. |
| Bitcoin | Yahoo: BTC-USD | Weekly level | −0.125 | High-beta liquidity/risk channel. |
| Gold | Yahoo: GC=F | Weekly level | −0.125 | Dollar/real-rate-sensitive store-of-value channel. |
| U.S. 2Y yield | FRED: DGS2 | Weekly yield level | +0.125 | Front-end rate and policy-expectations channel. |
| U.S. 10Y yield | FRED: DGS10 | Weekly yield level | +0.125 | Longer-rate and discount-rate channel. |
The signs are framework assumptions, not estimated regression coefficients. Relationships can change by regime. Equal nominal weights do not imply equal independent information or equal risk contribution.
Calibration period
January 1, 2015 onward
The production start date is 2015-01-01. The current model therefore does not provide a canonical 2008 observation.
Rebalancing
No weight rebalancing
The signed 12.5% weights are static. However, the full-sample mean and standard deviation are recomputed on each run as the sample expands.
Correlation
No explicit de-correlation
v2 does not apply PCA, covariance adjustment, risk parity or cluster caps. Correlated drivers can therefore reinforce one another.
Data treatment
Missing observations, freshness and outliers
Calendar alignment. After source provenance is captured, daily values may be forward-filled for at most three observations to bridge ordinary market-calendar mismatches.
Freshness gate. A release fails closed if a required driver is missing, future-dated or too stale. The production limits are 2 calendar days for Bitcoin, 3 for DXY/WTI/S&P 500/VIX/gold, and 4 for the two Treasury series.
Complete cases. After weekly resampling, any week still missing one of the eight required inputs is dropped from the calculation sample. The latest publishable week must be complete.
Outliers. Raw observations are not winsorized. Standardized component values are clipped only after z-scoring, at −3.5 and +3.5.
Source revisions. Yahoo/FRED histories can be revised or restated. Dated score archives preserve published score vintages, while the current full-history chart and robustness diagnostics are recalculated from the latest available source history. A separate revision audit measures the difference for each valid archived latest reading.
Regime labels
How score values are mapped to language
| Score range | Label |
|---|---|
| ≥ +1.0 | Strong dollar regime |
| +0.3 to < +1.0 | Firm dollar regime |
| −0.3 to < +0.3 | Neutral / transitional |
| −1.0 to < −0.3 | Soft dollar regime |
| < −1.0 | Weak dollar regime |
The regime thresholds are fixed specification choices. They are not estimated probabilities and should not be interpreted as confidence intervals.
Calibration and optimization
Was the model optimized in-sample?
The production code does not fit the eight weights or regime thresholds with a predictive optimizer. They are fixed constants chosen from the USD Impact transmission framework. That reduces one form of curve-fitting, but it does not eliminate model-selection bias: the choice of variables, signs, start date and thresholds remains discretionary.
The z-score normalization is explicitly full-sample. A new week changes the sample mean and standard deviation, which can revise historical z-scores and occasionally historical regime labels. For that reason, the recalculated historical series must not be presented as a point-in-time out-of-sample record.
Published validation evidence
What the robustness tests currently show
Material normalization sensitivity: across 556 evaluated weeks from January 2016 through August 21, 2026, expanding-window point-in-time normalization and the current full-sample recalculation produced the same regime label in only 40.1% of weeks. The mean absolute score difference was 0.512, with a maximum of 1.430. This is a robustness finding, not a predictive backtest.
Point-in-time normalization: published
Each evaluated week uses at least 52 prior complete weeks and normalization moments strictly before that week. Adding future observations is regression-tested not to alter earlier point-in-time scores.
Normalization choice materially changes history
Prior rolling windows of 104, 156 and 260 weeks showed only about 39.2%–43.4% regime-label agreement with the full-sample recalculation. Full-sample history should therefore not be treated as a stable point-in-time historical classification.
Correlation and contribution concentration are material
The latest 52-week rolling component diagnostic has U.S. 2Y/10Y correlation at 0.955. Absolute contribution shares imply 5.82 ordinary effective components, but the published absolute-correlation overlap heuristic reduces that to 1.89 effective correlated components — a 3.08× overlap multiplier. Gold is the largest absolute contributor at 25.25%. This is an audit/transparency diagnostic, not a covariance risk model or diversification estimate; v2 does not neutralize the overlap.
Driver and threshold sensitivity: published
Leave-one-driver-out regime-label agreement was about 83.4%–90.5% across the available sample. The latest reading stayed Soft dollar under seven single-driver omissions and became Weak dollar when the U.S. 10Y input was omitted. Narrower and wider regime thresholds retained about 85.4% and 90.3% historical label agreement with production thresholds.
True predictive out-of-sample test: preregistered, not started
A separate protocol was registered before its first eligible origin on August 28, 2026. It freezes the sign of the as-published Score v2 against the next completed-Friday DXY direction, requires 52 consecutive resolved predictions, prohibits backfill and interim performance reporting, and has no result yet.
Score v3 descriptive research: preregistered, not started
A separate prospective study freezes four candidate descriptive specifications before its first eligible week on August 28, 2026. It compares revision immunity, contribution concentration, leave-one-driver-out stability and regime turnover over 52 future weeks. It does not test predictive power, does not change production Score v2 and cannot automatically promote a candidate.
As-published revision audit: published
Each valid archive's declared latest observation is compared with the same week in the current recalculated history. Accepted source files are hashed, while invalid legacy archives are listed and excluded rather than repaired. The audit cannot separate expanding-sample normalization effects from upstream provider revisions.
2008: outside canonical v2
v2 begins in 2015 and includes Bitcoin. A 2008 result would require a separately labelled proxy study or a new methodology version; it cannot honestly be reported as a v2 observation.
The robustness and point-in-time files are current-vintage recalculations from the latest available provider histories. The separate vintage-comparison files use first-party as-published archives to audit revisions. Those completed studies are descriptive evidence, not independently audited performance or evidence of future predictive power. The prospective Score v2 predictive study and the separate Score v3 descriptive comparison are registered and automated but have not begun; neither supplies present performance evidence.
Reproducibility
Public artifacts
A reviewer can verify the current score arithmetic, inspect the production source, consume the machine-readable methodology contract, inspect the complete score history, review current point-in-time and robustness evidence, and compare valid as-published readings with current recalculations.
Production source codeExact calculation and data-treatment logic.Machine-readable methodology JSONProduction variables, sources, weights, normalization, clipping, missing-data rules, thresholds and scope boundaries.Methodology JSON SchemaClosed public schema for independent tooling and contract validation.Data semantics contractExact source fields, adjustment settings, holiday alignment, timestamp limits and futures-roll replication boundaries.Predictive-study preregistrationUntouched start date, one-week DXY target, fixed predictor, 52-outcome sample, comparators, inference calculation and success rule.Predictive implementation contractExact record sequence, source binding, tie handling, Wilson interval, circular shifts, benchmark lift and reporting gates.Pre-holdout engine lockSHA-256 hashes for the collector, evaluator, contracts and schemas frozen before the first eligible prediction.Score v3 descriptive-research preregistrationFour fixed candidates, untouched start date, prospective endpoints, eligibility gates, checkpoint rules and 52-week selection procedure.Score v3 metric contractExact definitions for the prospective integrity, concentration, stability and turnover endpoints used in the descriptive comparison.Score v3 pre-holdout engine lockSHA-256 hashes for the research engine, dependencies, protocol, metric contract and schemas frozen before the first prospective week.Latest decomposition JSONScore, driver z-scores, weights, contributions and source provenance.Full score JSONMetadata, weights, source provenance and weekly score history.Full score CSVWeekly component z-scores, score and regime.Robustness research JSONPoint-in-time summary, leave-one-out, sign sensitivity, rolling normalization, concentration, score distribution, regime duration, threshold and subperiod diagnostics.Point-in-time research JSONWeek-by-week expanding-window comparison with no future observations entering each week's normalization moments.As-published revision audit JSONHashed archive-by-archive score, regime and component revisions, plus explicit exclusions and limitations.As-published revision audit CSVTabular valid-vintage comparison for independent analysis and charting.Release reproducerStrict releases freeze the release week's levels, moments, z-scores, weights, contributions, provenance, code SHA and dependency-lock hash, plus per-driver and complete-matrix SHA-256 fingerprints at the daily and weekly stages.Source-retention policyProvider-rights boundary, evidence retained, explicit limitations and controls required before any raw-payload retention.Locked environmentExact production Python dependency versions.As-published archiveDated score and commentary vintages.
Strict-release evidence boundary. For releases generated from August 28, 2026, one same-run provider snapshot creates the score, a hashes-only daily receipt and the weekly input matrix. The bundle freezes the release week's exact levels, normalization moments, unclipped and clipped z-scores, weights, contributions, source provenance, pipeline commit and dependency-lock hash. It also freezes per-driver and complete-matrix SHA-256 fingerprints of the provider-derived daily histories after field selection and numeric parsing but before forward fill, plus the complete weekly input history. Original transport bytes are not hashed, provider-derived daily values are not published, and complete raw provider responses are not archived or publicly redistributed. The fingerprints can detect a later input-history change, but cannot independently reconstruct a changed provider history. Releases through August 21, 2026 predate this evidence bundle and remain legacy artifacts.
Institutional review questions
Direct answers
- What variables enter the model?
- DXY, WTI, S&P 500, VIX, Bitcoin, gold, U.S. 2-year yield and U.S. 10-year yield.
- What are the weights?
- Fixed magnitude 0.125 each; signs are +, −, −, +, −, −, +, + in the order above.
- How are variables normalized?
- Production uses full-sample z-scores of weekly levels using the sample mean and sample standard deviation, then clips each component to ±3.5. Separate published research tests prior-only expanding and rolling normalization.
- What is the historical calibration period?
- Production starts January 1, 2015 and expands weekly.
- How are correlations handled?
- They are not explicitly neutralized in v2. The latest 52-week rolling component diagnostic has a 0.955 U.S. 2Y/10Y correlation. A transparent absolute-correlation overlap heuristic maps 5.82 ordinary effective contribution components to 1.89 effective correlated components, a 3.08× overlap multiplier; this is an audit diagnostic, not a covariance risk model.
- How often are weights rebalanced?
- They are not. Weights are static; normalization moments update every run.
- How are missing observations treated?
- Limited daily forward fill for calendar alignment, strict freshness checks, then complete-case weekly filtering.
- How are outliers treated?
- Component z-scores are clipped at ±3.5; raw values are not winsorized.
- Predictive or descriptive?
- The production Score v2 is descriptive. A separate preregistered research protocol will test one narrow predictive question; it currently has zero resolved predictions and authorizes no predictive claim.
- Optimized in-sample?
- No predictive optimizer fits the production weights or thresholds, but the specification choices remain discretionary.
- Backtested out-of-sample?
- No completed predictive out-of-sample result exists. A future-only one-week DXY study is preregistered to begin August 28, 2026 and requires 52 resolved predictions; the earlier point-in-time normalization study remains a descriptive robustness test.
- Stable across regimes?
- Not invariant. The published robustness battery finds 40.1% regime-label agreement between point-in-time and full-sample normalization across 556 evaluated weeks, while leave-one-driver-out agreement is materially higher at about 83.4%–90.5%.
- 2008, 2020, 2022?
- 2008 is outside canonical v2. The published research includes 2020 and 2022 anchor-window comparisons under both point-in-time and full-sample normalization, but these are descriptive robustness checks rather than predictive tests.