Hull-White’s Mean-Reversion Problem: A ‘Smarter’ Model Lost to a Naive Guess in 2022
Mean reversion recalibrates monthly. Volatility recalibrates daily. Here’s why that gap matters — and why it isn’t a trading edge yet.
In every one-factor Hull-White implementation, dr = (θ(t) − ar)dt + σdW, mean reversion and volatility get calibrated on different clocks: volatility daily, mean reversion roughly monthly. That cadence gap is real, documented independently by a bank’s own quant team and by academic researchers, and there’s a defensible reason for it. What’s far less established is what follows from it. Exactly one vendor, one currency, and one six-month window has ever shown this gap producing a measurable valuation error — and nobody, including this piece, has shown that error converts into money changing hands. What follows separates the mechanism, which holds up, from the pattern claim, which doesn’t yet.
The consensus, and the parameter underneath it
Practitioners use Hull-White because letting θ(t) absorb the shape of today’s discount curve preserves closed-form bond and swaption prices while still fitting the market exactly (Hull and White, NYU Stern working paper). The well-known limitation is that a single Brownian driver forces every point on the curve to move in lockstep, which risk-validation teams flag as the model’s central weakness (RiskSpan). That limitation is precisely why Longstaff, Santa-Clara, and Schwartz’s 2001 study of American-style swaption exercise found that, based on ISDA notional estimates, following a myopic single-factor exercise strategy rather than a multi-factor one cost swaption holders on the order of several billion dollars in aggregate (Longstaff, Santa-Clara, and Schwartz, Journal of Financial Economics, via SSRN). None of that is new to anyone trading these books. The less-examined mechanism sits inside the one-factor model itself, in how mean reversion, a, is actually estimated day to day.
The math, stated plainly
Bond prices under Hull-White are exponential-affine
and θ(t) is solved so the model reproduces the observed instantaneous forward curve exactly:
(Brigo and Mercurio’s derivation, reproduced in “Consistency of extended Nelson-Siegel curve families with the Ho-Lee and Hull and White short rate models”). B(t,T) is the function that matters here: it’s the exponential-decay term determining how a shock to the short rate propagates along the curve, which is why a — not σ — controls the shape of the model-implied volatility surface along the tenor axis, while σ governs its shape along the expiry axis (S&P Global Market Intelligence, 2023).
Two clocks, and what’s actually corroborated
Here is the part with genuine independent support. S&P Global’s research, produced by the quant team running a bank’s own CVA/XVA calibration, states that a mean-reversion update cadence of roughly once a month, against a daily update for volatility, is what they find reasonable in practice. A separate, peer-reviewed 2025 paper in Quantitative Finance reaches the same operational conclusion from a different angle: because mean reversion and volatility have overlapping effects on Black implied vols within a single expiry-tenor smile, a is typically held constant and re-calibrated only weekly or monthly, “due to the limited impact of the mean-reversion parameter, combined with the absence of market data to calibrate this parameter to,” while volatility is re-calibrated daily (van der Zwaard, Grzelak, and Oosterlee, Quantitative Finance, 2025). Two independent teams, different institutions, agreeing on the practice itself. That’s as far as the corroboration goes, and it’s worth being precise about that boundary: neither paper, nor any other source in this piece, independently confirms any of the specific numbers in the next two sections. Those come from one place.
Where the mechanism bites hardest
This cadence gap matters disproportionately for anything with early-exercise features, because of how B(t,T) works. A Bermudan swaption’s Vega is distributed across the entire term/tenor structure rather than concentrated at one point: a representative 11-non-call-1 Bermudan receiver has a 1bp Vega of 0.06% of notional, against Vegas on its co-terminal European swaptions ranging from 0.009% to 0.048% (Gatarek and Jabłecki, Mathematics, 2021). Since B(t,T) is exactly the function mean reversion controls, a stale a misprices a Bermudan more than it misprices any single vanilla instrument used to calibrate it — and that exposure is not a corner case: roughly 60% of bonds in the Bloomberg Barclays Global Aggregate Credit Index carry call provisions that get stripped and sold on as swaptions (Gatarek and Jabłecki, 2021). For non-standard amortizing or accreting Bermudans specifically, the problem compounds further: their own calibration targets are “generally no more liquid than the [instrument] itself,” so even the calibration inputs have to be inferred rather than observed (Risk.net, “Cutting edge: Option pricing — Bounding Bermudans”). This is the strongest part of the argument, and it’s a statement about exposure, not about profit.
The one data point that exists
Benchmarking a book of at-the-money swaps out to 30 years against exact market-implied CVA, one vendor’s fully calibrated Hull-White model produced a CVA root-mean-square relative error (RMSRE) “consistently of the order of 5% or less” — except in the first half of 2022, when the euro book’s error briefly breached that threshold. Freezing mean reversion at a naive constant 5%/year instead, and recalibrating only volatility, kept CVA RMSRE largely below 10% and, in that same window, actually beat the fully optimized model (S&P Global Market Intelligence, 2023). A separate 2020 study from the same research team found optimized mean reversion for EUR, JPY, and USD all crossing into negative territory that spring — a value with no coherent identity as a “speed of reversion” (Puetter and Renzitti, IHS Markit, 2020).
That is the entire empirical record. It is one currency book (EUR), one implementation, one vendor’s research team, across one six-month window. It has not been shown to hold for USD or JPY in the same period, and it has not been tested at all against the other monetary-policy inflections of the last fifteen years — the 2015 Fed liftoff, the 2018 hiking cycle, the 2020 cutting cycle. A pattern observed exactly once, by the same team whose own methodology produced it, is a hypothesis worth taking seriously — not yet a structural, recurring feature of the market.
The gap between mispriced and profitable
Every number above is a valuation-accuracy metric: how far a model’s output sat from a market-implied benchmark. None of it is a trade, a position, or realized P&L. Showing that a bank’s own optimized calibration was less accurate than a naive constant for six months establishes that a valuation gap existed inside one firm’s risk system — it does not establish that a counterparty on the other side of a trade was pricing off the worse convention, that the gap was large enough after transaction costs and bid-offer to act on, or that anyone actually captured it. Bridging “our model disagreed with theirs” to “here is how you extract money from that disagreement” requires evidence this piece does not have: a documented instance of a specific mispriced trade, a counterparty using a demonstrably stale convention, or a realized P&L outcome tied to this specific mechanism. Absent that, the honest claim is that a real, mechanism-grounded valuation discrepancy exists and has been observed once — not that it is a capacity-bound edge with a known scale.
What would actually confirm or kill this
Confirmation requires three things this piece doesn’t have: the same RMSRE reversal appearing in USD or JPY books during the same 2022 window (testing whether it’s currency-specific or general); the same pattern appearing around 2015, 2018, or 2020 inflections (testing whether it’s a recurring feature or a one-time artifact of 2022 specifically); and a second bank’s independently built CVA system showing the same result (testing whether it’s a property of the market or an artifact of one vendor’s implementation). Any of these failing to replicate would be reason to treat the 2022 finding as an isolated event rather than a mechanism-driven pattern. None of these tests has been run in the public literature as far as this piece can establish.
What to do with this today
The mechanism — mean reversion calibrated on a slower, less-disciplined clock than volatility, with disproportionate impact on Bermudan and amortizing structures — is well enough established to be worth a risk manager’s attention as a diagnostic. A fitted mean reversion that has drifted toward zero or turned negative around a policy inflection is worth treating as a signal that the model is curve-fitting through a regime shift rather than reading it, and a CVA or exposure model on a Bermudan-heavy book is worth benchmarking against a naive fixed-mean-reversion alternative at FOMC and ECB inflection points, simply because the one time this has been tested, the naive benchmark won. That is a reasonable, low-cost risk-management practice, and it is a different and smaller claim than “this is a tradeable edge” — the evidence here supports the former, not yet the latter.
📊 Want Deeper Quantitative Analysis?
This research took a significant amount of time — data collection, verification, and analysis. If you found value in this deep-dive, I publish exclusive quantitative research, trading strategies, and institutional-grade analysis on Patreon.
By joining, you’ll be supporting my work and motivating me to publish more content like this.
→ Join the Patreon community here
Connect with me: