Memories of the LQG 2025/09/03 – 2025/09/06 – Autumn Seminar – at Cambridge

A quant remembers … the Robert Macrae LQG Journal 2025

While some might journal holidays, expeditions, family events and so on, the mark of a real quant is the choice of topic to write up and keep.

Below is a fine example of both topic choice and journaling accuracy and style.

“The LQG autumn seminar was, as always, graced by a range of fine presentations.
With LQG’s thanks to our speakers and my apologies for the inevitable misunderstandings and omissions.”  Robert Macrae

David Buckle, Inquire
E[g(X_1, …, X_n)] = ∫[0,1] … ∫[0,1] g(F_1^(-1)(u_1), …, F_n^(-1)(u_n)) . c(u_1,…,u_n) du_1 … du_n  for dummies

David explained how to sample from a kernel-smoothed 1D empirical distribution.  For the 1% tails he assumed exponentials chosen to match empirical skew and kurtosis, resulting in blindingly fast run-time typically requiring 1 linear interpolation occasionally one log.

Armed with this he approached multivariate sampling via Gaussian copulas, using a root-finding algorithm to ensure that observed correlations are preserved as in general these differ from the copula correlations.  So conditional on tail co-dependence and other requisite copula assumptions you can sample from the prospective returns of any portfolio, and hence directly optimise any utility function based on returns.  You can also scale and shift those returns to investigate how different assumptions should affect choices, for example to investigate the impact of house strategic views.

Naima Hammoud, Bloomberg
Extracting transient factors from cross-sectional residuals

Big market moves (in particular) induce correlations in idiosyncratic returns to a pre-specified factor model.  Naima pointed out that the RMS T-statistic of a factor K can be interpreted as bias to risk if K omitted and shows that performance is significantly improved by incorporating the principal component of the residuals as an additional factor.  105 days is a sensible estimation half-life and shows interesting dynamics around large events such as Covid, possibly picking up an AI factor.  Simon questioned whether weighting by inverse specific variance, despite its elegance, necessarily improved performance for near-cap-weighted portfolios.

David Aikman, NIESR
In the Fed we trust?

Picking a fight with the Chancellor in order to maximise the newsworthiness of an LQG talk sets a high bar for future inaugural speakers.

Money is a social convention underwritten by trust, and trust in central banks should increase resilience to political interference.  To measure trust David uses tweets.  While substantial problems are involved — tribal and bot activity, and according to Ely “People with nothing better to do” — they can sensibly consider covariates and use scandal-related resignations as a plausibly-exogenous instrument.

Fine tuning proved unsatisfactory so they just asked ChatGPT-4o mini to categorise 4 million central-bank-related tweets and retweets as positive, neutral or negative on trust and accumulated the smoothed result as their trust index. This suggests a plausible pattern with falling trust under Trump 1; partial recovery under Covid; and a further fall under inflation.  Given the smooth nature of the index the regression and VAR analyses appeared interesting rather than compelling.

Jose Menchero, Bloomberg
Factor misalignment in portfolio construction

To investigate the quality of a risk model you should compare realised to forecast volatility on hedged portfolios because a long only constraint regularises and can hide problems.  Jose removed each factor in turn from his prespecified-factor covariance model and instead treated the factor exposures as alphas to create portfolios highly exposed to the missing factor and so give a sensitive test of the value of including that factor.  He concluded that gains from inclusion were on the scale of 6% reduction in realised volatility, which is comparable to the estimated gain from using inverse specfic variance weights over sqrt(market cap).   By contrast there seems no value in including alphas to the extent that they are orthogonal to all factors.

Annika Rao-Monari, CleeAI
From syntactic to semantic AI

Chatbots know what an answer looks like and slot material into this structure as appears most appropriate, playing Tetris with sentences.  Syntax is king and while results can be impressive, LLMs hallucinate when they cannot find a reality-based sentence that is a good fit to the emerging pattern.  Annika argued that to progress further we may need a more evaluative approach centred on concepts rather than sentences.

James Eagle
Data visualisation

With Trump firing Fed governors and a labour statistics commissioners, can we assume US institutional reports will remain accurate?  While outright fabrication may remain rare there is endless scope for the manipulation of methodology and presentation.  The less we can trust institutions the more effort we will have to invest in verification and interpretation, and the greater the need for trustworthy, effective and digestible data-based narratives.  James illustrated with a number of examples.

Desi Ivanova, Oxford
Complexity, uncertainty and learning

Desi gave a thorough analysis of the Virtues of Complexity and double descent, showing that the modern ML incarnation is closely related to the familiar pseudoinverse effect in which underspecified regressions are fitted with minimum norm.   She also showed that the influential Kelly paper implied that radial basis functions would perform even better on their problem than their preferred random fourier factor approach, and with vastly less computation.

More generally, Desi argued for a change in approach away from explicitly parsimonious models towards large, rich, complex, non-convex models estimated via gradient descent and with overfitting controlled by careful use of holdouts, dropouts and other modern approaches.  Overfitting should only be a problems when those models are tested outside the domain in which they have been trained.   Old-school quants may well agree… but, bruised by long experience, also suspect that the future holds many such places…

Dori Levanoni, Capstone
Optimal currency hedging

International investments involve exposure to both an asset and a currency and Dori asked how investors might hedge the latter.  Since 1974 from the perspective of a US or UK investor the currency exposure appears to have a small positive association with equity assets but also a small positive return.  Of course for Japanese investors the situation appears very different.  While the USD has been widely discussed as a safe haven since GFC this was far less prevalent in the decades following the collapse of Bretton Woods.  Dori considered a trading strategy that implemented whatever hedge would have performed best during the previous 3 years and found some marginally LQG-compliant evidence that this had performed well.

Pedro Castro, Man Group
FX hedging

Currency volatility is on the scale of 11%, too large to be ignored, and institutions typically hedge about half their FX exposure.  This roughly corresponds to the historic mean-variance optimising equity hedge over 50 years, but outcomes across markets and decades show huge dispersion.  Pedro presented some evidence that hedging so as to maximise carry may perform better than a fixed strategy, though in developed markets the gains are small and mainly accrue before GFC. There may be a small additional improvement from penalising recent volatility, while momentum and PPP seem to add little.

Simon Polbennikov, Barclays
Bond performance attribution

RAIL — return attribution at issue level — starts with a bond’s excess return over duration-matched treasuries and decomposes as spread carry; systematic; and specific components. Systematic contains market; spread duration; quality; and sector components.  All are scaled by DTS (Duration * Spread) ratios.  These systematic exposures account for a substantial part of bond volatility and removing them permits more accurate attribution of performance.  For example, Simon estimated that analyst recommendations generate around 0.15% over the following 6 months.

Ed Fishwick, Blackrock
Themes repeat

Some LQG luminaries  — the Vivaldis of finance — recycle remarkably similar material for decades, whether reverse optimisation or meditations on Chapter 12 uncertainties.

Ed noted that smoothing in PE and RE markets rendered them unrealistically attractive compared to public equities in terms of volatility-based metrics.  However, he questioned which valuation was more appropriate.  In the 1700s when packet boats could not sail between Amsterdam and London the volatility of dual-listed stocks was little changed despite the absence of news.  Should we conclude that excess volatility is intrinsic to public markets; that short-term volatility is a poor risk measure; that public markets should converge towards private market valuations; and that  “The rest is noise”?  This last is apparently a tribute to Alex Ross’ magisterial review of 20th century classical music.