The 551 bp Treasury reset.
Macrovision replayed the rise in the 3-month Treasury yield from 0.08% to 5.59%, rejected a weaker trend-chasing model at every tested horizon, and traced a current market-forward rate into bill value. The case proves calculation lineage, model selection and suppression discipline. The case does not prove that Macrovision predicted the 2022 reset in real time, caused an outcome or delivered realised investment performance.
What the case proves
The case proves that Macrovision can transform dated public observations into a point-in-time panel, run competing formulas on common data, reject an inferior candidate, select an eligible rate path, and translate that path into deterministic Treasury-bill maths. Every stated run carries its method version, source date and calculation hash where the current contract exposes one.
The case does not prove prospective forecast skill or financial alpha. A prospective performance claim requires timestamped forecasts made before outcomes, non-overlapping or overlap-adjusted evidence, uncertainty intervals, costs and a defined counterfactual. Macrovision has the recorder for that evidence, but a statistically meaningful matured live history is not yet available.
Example 1 · The 2022–23 Treasury reset
The 3-month Treasury constant-maturity yield rose from 0.08% on 3 January 2022 to 5.59% on 31 October 2023, a 551 bp move. The series reached 5.63% on 6 October 2023. Macrovision replayed 458 official FRED observations through the current persistence and five-day momentum formulas.
| 1 business day | Persistence 3.3895 bp MAE · momentum 3.6522 bp · persistence lower by 0.2627 bp |
| 5 business days | Persistence 8.2274 bp MAE · momentum 9.2082 bp · persistence lower by 0.9808 bp |
| 21 business days | Persistence 27.2128 bp MAE · momentum 32.5652 bp · persistence lower by 5.3524 bp |
The replay maths
- Persistence.
predicted change(t,h) = 0.- Five-day momentum.
predicted change(t,h) = h × (yield(t) − yield(t−5)) / 5.- Mean absolute error.
MAE(bp) = 100 × mean(|predicted change − realised change|).- Root mean square error.
RMSE(bp) = 100 × √mean((predicted change − realised change)²).- Bias.
bias(bp) = 100 × mean(predicted change − realised change).
Value created: the gate prevented the more active but less accurate momentum rule from being treated as an improvement. The demonstrated value is disciplined model rejection, not a claim that a cash return was earned.
Example 2 · A current market-forward projection
A separate FRED-backed run as of 15 July 2026 started with a 3-month yield of 3.83% and a 6-month companion yield of 3.93%. Macrovision solved the same-curve ACT/365 simple-interest identity for a 3.991882% forward rate.
(1 + y6 × 182/365) = (1 + y3 × 91/365) × (1 + f × 91/365)
f = (((1 + 0.0393 × 182/365) / (1 + 0.0383 × 91/365)) − 1) × 365/91 = 3.991882%.
The publication tournament
| Forward | 21.9969 bp MAE · 26.6877 bp RMSE · selected |
| Persistence | 36.0000 bp MAE · 42.8395 bp RMSE |
| Structural | 43.6464 bp MAE · 54.5072 bp RMSE · optimistic replay limitation |
| Momentum | 57.9683 bp MAE · 77.4027 bp RMSE |
| Ridge | 59.0068 bp MAE · 68.8382 bp RMSE |
published path = arg min eligible model(latest walk-forward MAE), with at least 20 observations and ties resolved in favour of fewer estimated parameters. The forward MAE was (36.0000 − 21.9969) / 36.0000 = 38.9% lower than persistence in this replay.
The comparison uses the latest 252 overlapping folds. The folds are not 252 independent observations, and the structural score applies today’s macro impulses to historical folds. The comparison is research evidence for selecting the stated path, not statistical proof of repeatable forecast skill.
From yield to bill value
- Bill price.
price = face / (1 + bond-equivalent yield × days to maturity / 365).- Rolling ladder.
face(k) = value(k) / price(k); maturity proceeds become the next bill’s value until the horizon.- Value identity.
final value = initial cash + carry contribution + repricing contribution.
For modelled initial cash of $1,000,000 over 126 business days, equal to 176 calendar days in this run, the selected path produced $1,000,000 + $18,927.543235 + $0 = $1,018,927.543235, or 1.892754%. The amount is a projection from stated rates and conventions, not an actual, promised or guaranteed return.
Example 3 · The live public Macrovision snapshot
The public Lens snapshot captured on 17 July 2026 shows the broader Macrovision formula pack working on the same governed data estate. The United States readings were PMRS 63.00, real-economy resilience 73.67, fragility pressure 49.75, macro-risk transmission divergence 23.92 and Lens confidence 74.68. The global Treasury Liquidity Absorption Index was 42.88.
| Public Macro Regime Score | 63.00 · expansionary |
| Real-Economy Resilience | 73.67 · resilient |
| Fragility Pressure Score | 49.75 · contained pressure |
| Macro-Risk Transmission Divergence | 23.92 · late-cycle tension |
| Lens confidence | 74.68 · good confidence |
| Treasury Liquidity Absorption | 42.88 · neutral |
The snapshot published explicit suppressed inputs and one input hash, method hash, output hash and calculation hash. Those hashes prove reproducible lineage for the captured inputs and method; hashes do not prove economic truth or forecast accuracy. The Lens formula-pack readings are presentation analytics and did not feed the selected market-forward path in Example 2.
The controlled dependency graph
Macrovision uses a one-way, gated dependency graph. The graph makes every active and inactive branch visible, rather than allowing every measure to influence every output.
0.35 prices + 0.15 labour + 0.15 activity + 0.15 money + 0.10 sovereign + 0.10 relative bias. Promoted risk edges may adjust this branch.The implemented maths across the full spine
- Feature change.
percentage change = (current / previous − 1) × 100;annualised change = ((current / previous)^periods − 1) × 100.- Shared score rail.
score = clamp((metric − reference) / band × 10, −10, +10), with the registered direction applied before rounding.- Regional headline.
headline = Σ(pillar score × weight) / Σ(covered weight), followed by the registered imbalance adjustment and a clamp to[−10,+10].- Regional confidence.
0.30 coverage + 0.20 freshness + 0.20 revision quality + 0.20 coherence + 0.10 duration.- United States–Euro Area bias.
(US regional signal − EA regional signal) × min(US confidence, EA confidence).- Structural Treasury candidate.
0.55 momentum + 0.35 macro impulse × √progress + 0.08 curve slope × √progress + 0.02 FX bias × √progress. The macro impulse isclamp(policy pressure / 10, −1.5,+1.5).- Wrapper model blend.
model weight ∝ exp(−η × loss);shrinkage weight = n × quality / (n × quality + κ);shrunk APY = weight × model APY + (1 − weight) × peer APY.- Wrapper uncertainty and daily return.
uncertainty = √(variance + (1 − quality) × ambiguity²);daily return = (1 + APY/100)^(1/365) − 1.- Allocation and decision gate.
utility = expected edge − tail − uncertainty − proof − liquidity − operational − turnover − concentration penalties. A MOVE also has to pass probability, expected-edge, fifth-percentile downside and breakeven tests against the greater of 5 bp or the uncertainty-plus-operational buffer.
The formulas above describe implemented layers, not one claim that every layer was active. The historical replay stopped at model evaluation, the current projection stopped at bill value, and no wrapper allocation or MOVE decision was evaluated in this case.
Does everything feed everything?
No, and that is the correct design. Uncontrolled cross-feeding would introduce circularity, double-counting and unstable feedback. The defensible Macrovision capability is that every output is traceable to explicit inputs and formulas, while only eligible, validated and available edges feed a downstream calculation.
The current FRED-only value run had no database-backed regional pillars and no risk-transmission surface. The winning forward model did not use those branches. Macrovision exposed the branches as unavailable or inactive instead of pretending they contributed. That suppression is part of the proof.
How the capability creates commercial value
- Free Lens.
- Public-source macro, Treasury-rate, risk, freshness and method context makes the capability inspectable and creates a credible entry point.
- Manually gated Intelligence and Pro.
- Professional users can review deeper scorecards, uncertainty, model comparisons, historical context and the developing prospective track record.
- Governed institutional engagements.
- A team can test its own policy and evidence assumptions without YieldGuard taking custody, routing assets or making a discretionary decision, then retain a reviewable decision packet.
- Revenue boundary.
- Commercial value comes from access, workflow, evidence and usage. YieldGuard does not claim a share of forecast alpha or promise a financial return.
Evidence, versions and sources
- Historical replay.
- 3 January 2022 to 31 October 2023 · 458 DGS3MO rows · panel hash
c1fbd7a69a8f8b62.e7b4dcf4525e4ada 5a57af2ae04d9d2e 18f562e21fc7d501 - Current research panel.
- 1 September 1981 to 15 July 2026 · 11,216 rows · 100% FEDFUNDS and DGS2 coverage · panel hash
81651d9961bc0877.4eb0fd3176e61458 253ee874bf1ed521 1be2fa9d87f87159 - Method versions.
treasury-research-baselines-v2,treasury-bill-value-v3,treasury-forward-implied-v1andlens-formula-pack-v0.4.- Primary public sources.
- FRED DGS3MO, FRED DGS6MO, FRED DGS2 and FRED FEDFUNDS. Source organisations do not endorse YieldGuard or Macrovision.
Questions
Did Macrovision predict the 2022 Treasury reset in advance?
No. The historical example replays the current method over public observations. The example demonstrates model comparison and rejection, not a real-time prediction made in 2022.
Does every formula feed the selected Treasury projection?
No. Only explicit, eligible and available dependencies feed an output. In the current case the forward model won, so the macro and risk branches did not alter the selected path.
What value does the case demonstrate?
The case demonstrates decision-support value: a weaker model was rejected, inactive inputs were exposed, calculations were reproducible and the selected rate path was translated into reviewable bill-value maths.
Is the projected Treasury value an actual or guaranteed return?
No. The figure is a modelled projection using stated public rates, day-count conventions and assumptions. The figure is not an actual return, recommendation or guarantee.
Compliance boundary
YieldGuard Ltd, company number 16914415, registered in England and Wales, is a technology provider. YieldGuard does not hold assets or private keys, provide investment advice, act as broker or dealer, route or execute transactions, or make discretionary investment decisions. Every decision remains with the professional user, and no outcome or return is guaranteed.
Institutional service enquiry