Architecture, risk, and validation of the Vanttax strategies

What problem the two strategies solve

A daily-reset leveraged ETF is not a long-term leveraged bet: it is an instrument that resets its exposure every session. Holding it passively exposes the investor to a mathematical drag (developed in section 6) that erodes results in sideways or volatile markets, even when the underlying direction is correct. Both strategies are designed on that premise: leverage is only defensible when paired with an explicit rule for when to hold it and when to move to cash.

Monetary Hedge and TQQQ Momentum solve that problem from deliberately decorrelated angles. Monetary Hedge rotates among a small basket of leveraged assets — equities, long-dated sovereign bonds, and gold — according to the macro and volatility regime its model identifies, with a structural tilt toward sovereign debt that, as explained in section 2, is not hard-coded but emerges from which asset has consistently offered the best risk-adjusted profile over four decades of data. TQQQ Momentum is narrower by design: it operates a single asset (leveraged tech index, or its UCITS equivalent in Europe) versus cash, and its decision loop is built specifically around that instrument's behavior.

Neither strategy promises to be always invested. Cash is a legitimate position, not an absence of strategy.

Decision architecture

Each strategy relies on a supervised learning model (gradient-boosted decision trees combined with a trend-following component as fallback when primary model confidence is low) trained on feature families with explicit economic rationale, not thousands of brute-force-mined technical indicators.

Monetary Hedge

The model evaluates, day by day, which regime the market is in by combining five information blocks: price trend and momentum across horizons; volatility regime (level, term structure, and historical VIX percentile); macro context (money supply, yield curve, real rates, credit spreads); fundamental valuation (Shiller CAPE vs. long-run mean); and cross-asset momentum. On top of model output, an additional risk-management rule layer — independent of machine learning — can force cash or reduce leverage when stress signals of different types coincide.

TQQQ Momentum

This is a binary classifier focused on a single asset, with variables specific to that instrument: multi-horizon momentum, position vs. long moving averages, implied volatility level and recent change, drawdown from highs, and yield-curve slope. A relevant design choice: the risk-adjusted return threshold defining the training label is not a fixed number — it is a dynamic percentile computed over the historical distribution observed up to that point. This avoids anchoring the model to an arbitrary value chosen in hindsight because it "worked well" in the past.

Risk management and execution

No position opens at market without confirmation. Entry requires price to break a reference level above the prior high or close, with a bounded waiting window before a fallback entry — reducing exposure to signals that reverse the next day. Once in a position, each trade carries a dynamic stop based on that asset's recent volatility (wider for more leveraged instruments, tighter for less volatile ones), bounded between minimum and maximum, with a minimum holding period to avoid noise-driven churn.

Position sizing is binary: full exposure to the chosen asset or full cash — never a fractional blend. Cash is not idle — it earns the reference overnight rate of each market. Commissions and execution slippage are modeled on every entry and exit, not only in aggregate results.

Purged walk-forward validation

The system is not validated with a single train/test split. It trains on a multi-year window, drops a purge gap between the end of training and the start of testing — to remove leakage when a label computed weeks ahead "touches" the test period without that gap — and evaluates exclusively on the subsequent segment never seen by that model fit. The window rolls forward and the process repeats across the full available history, producing dozens of independent out-of-sample results spanning the dot-com bust, 2008 financial crisis, 2020 shock, and 2022 rate-hike cycle.

A governance point we consider relevant for ruling out overfitting: model configuration (tree depth, learning rate, regularization) is fixed — the same in every segment and in daily production inference. Only the amount of history used for retraining grows over time, never the recipe itself. Re-optimizing configuration against recent performance would be precisely the form of overfitting this process is designed to prevent. Production retraining is anchored to the horizon that defines the model label — several weeks of market time — not an arbitrarily chosen calendar.

Eight-test anti-overfitting battery

Before a configuration is considered fit to operate, it must pass eight independent statistical tests. None replaces the others — they were designed to attack overfitting from different angles:

All eight tests, with results for each model, are generated automatically on every validation run — not a one-off manual review.

The mathematics of daily leverage

A daily-reset leveraged product does not multiply underlying return over time — only session by session — and that daily reset introduces mathematically inevitable friction, a direct consequence of applying Itô's lemma to daily return compounding:

The practical consequence is that a 3× product can return less than three times the underlying even when direction is correct, if the path is volatile. This is not a quirk of one product: it is a structural property of any daily-reset leveraged instrument, and is the fundamental reason this document advocates active exposure management — with explicit entry, exit, and stop rules — over passive multi-year holds.

Data integrity and provenance

Market prices come from standard public sources, dividend- and split-adjusted. Macro series are sourced from the Federal Reserve (FRED) with a fallback chain so we do not depend on a single vendor, plus Shiller CAPE fundamental valuation. A methodological detail we consider uncommon in retail quantitative analysis: macro series are time-shifted according to their real publication lag (inflation and employment data are not publicly available until weeks after the period they refer to) before entering the model, so the system never "sees" macro data before it actually existed in the world.

Part of the history backing reported results corresponds to reconstructed prices, not live market quotes. Several leveraged instruments used — especially European UCITS products — have recent listing dates relative to the analyzed period. For those segments, price is rebuilt from the underlying asset or index using the same daily-reset formula in section 6, then spliced to the fund's live price from its listing date, rescaling at the join to avoid an artificial jump. This simulated segment is explicitly marked in every performance report — it is not presented as a live traded track record.

Independent execution by market

The regime decision (which asset type to hold) is computed once and shared between the US and European versions of each strategy. What is not shared is execution: entry confirmation and dynamic stop levels are computed independently for each instrument, on its own price history, with risk parameters adjusted to its leverage and volatility — a 5× UCITS product does not have the same risk profile as its 3× US equivalent.

The observable consequence is that, in specific segments, one portfolio and the other may not match exactly on whether they are in a position, even sharing the same underlying regime diagnosis. That is not system inconsistency: it is the expected result of modeling execution on each instrument's real behavior, rather than assuming both markets move identically.

Institutional precedent for managed leverage

Applying leverage to a risk-managed portfolio, rather than concentrating unlevered capital in the highest expected-return asset, has a long track record in institutional management — not an argument invented to justify a retail product:

• "Why Not 100% Equities" — the core argument is to find the best return/risk portfolio first, then adjust risk with leverage, instead of concentrating capital in the most volatile asset. AQR — A. Ilmanen
• "Risk Parity: Why We Lever" — explicit institutional justification for leverage on diversified low-volatility portfolios. AQR
• All Weather (1996) — pioneering risk-parity strategy, leverage on a risk-balanced portfolio to reach a return target. Institutional-only for years. Bridgewater Associates
• Kelly criterion, academic formalization of optimal leverage given measurable edge, applied by E. Thorp at Princeton Newport Partners (1969–1988). J. L. Kelly / E. Thorp

These references document that the principle — not the result of any specific product — is recognized institutional practice.

Disclosures and limits of this document