CV, research and teaching

Career and research

“The machine calculates. The people who decide should understand why.”

How do a family’s goals become a portfolio that lasts a lifetime? I look for answers with data, statistics and artificial intelligence.

Download CV (PDF)

Timeline

Since 2013 I have done research on investing, and since 2016 also together with students: in 32 master’s theses, industry projects and internships with more than 50 students from six universities I was closely involved from the question to the result. Add four collaborations with doctoral researchers and institutes, and my lectures. And at work I do research every day, since 2008 at Sal. Oppenheim, HQ Asset Management and FINVIA; the row Work shows when. The figure shows where each project belongs in the investment process.

Projects and teaching by stage and year

Work

Student jobs FINVIA HQ Asset Management Sal. Oppenheim, from 2008

Education

PhD, LMU Munich Diploma

with studentswith co-authors, doctoral researchers and instituteslecturesummer schoolseveral years One mark per project or lecture, at the year it started; a line shows how long it ran, an arrow that it is still running. The rows Work and Education show where I worked and studied; the diamonds are summer schools. The grey column on the left sums up 2004 to 2011. Titles in the original language.

2012Lecture, Koblenz University of Applied SciencesMethods and teaching
Financial Innovations (MSc)
2013–2019Working paper, University of CologneExpectations
Tree-based conditional portfolio sorts
Tom Zimmermann
Decision trees and random forests show which of the many indicators really explain future stock returns.
Best Paper Award 2015
To the paper
2016–2020Chapter of the PhD thesis, LMU MunichEnvironment
The relation between stock market risk and return
Stefan Mittnik
2016Master’s thesis, University of ZurichGoals
Assessing the risk tolerance of retail investors: risk profiling in theory and practice
Claudio Ambühl
Online risk questionnaires are not reliable: the same test client was recommended anything from 23 to 90 per cent in shares.
Same client, 21 online questionnaires: how much in shares?

rated 5 out of 10rated differently

Test client: 30 years old, no prior knowledge, wants to invest 15,000 for 15 years. One bar per online provider: the share of shares in the recommended mix, rounded. Even those that rated him the same (dark) recommended between 23 and 61 per cent in shares. Source: master’s thesis of Claudio Ambühl, University of Zurich, table 8; own chart.
- Eight of the 21 providers recommended the same portfolio to all three test clients: to the 30-year-old as well as to the 60-year-old who retires in six years.
- Only about four in ten questionnaires met the basic EU rule for investment advice (MiFID), and none asked separately how much risk someone wants to take and how much they can bear.
2018PhD thesis, advisor Stefan Mittnik, LMU MunichExpectations
Applications of textual analysis and machine learning in asset pricing
PLEXUS Advancement Award 2020
Thesis (PDF)
2018Industry project, LMU MunichEnvironment
Finanzmarktprognose mit Hilfe von Textdaten
Karem El-Oraby and Raphael Rehms
Articles from DIE ZEIT make monthly forecasts for German shares, bonds and the Ifo index slightly better: the forecast error falls by up to 9 per cent.
How much smaller the forecast error became with newspaper text

Euro Bund future8.7%
Ifo business climate6.5%
VDAX (swings of the DAX)3.1%
DAX2.6%
REX bond index0.3%

Forecasts for the following month, on the last third of the data up to November 2017, each made only with what was known at the time. Compared with a model that uses only the past of the series itself; for each series the best of 220 variants, so rather an upper bound. Source: project report of Karem El-Oraby and Raphael Rehms, LMU Munich, 2018, section 5.3; own chart.
- The basis: 469,137 articles of DIE ZEIT since 1946, turned into word counts month by month. A statistical method (LASSO) picked, out of 225,528 words, the few that helped the forecast.
- Many of the words chosen had nothing to do with markets in terms of content: a reminder of how easily chance patterns turn up in large amounts of data.
2018Master’s thesis, LMU MunichExpectations
Forecasting financial time series with deep learning
David Milewski
Neural networks called the direction of the US stock market right in up to 74 of 100 months; always saying “up” was right in 69.
How often the direction was right: 108 months of US shares

Always saying “up”68.5%
Classic time-series model (ARIMA)66.7%
648 neural networks (LSTM)47 to 74%

Share of months in which the forecast for the following month had the right sign, November 2008 to October 2017, US total market index; the networks were trained on monthly returns from 1959 to 2008. Source: master’s thesis of David Milewski, LMU Munich, 2018, section 3.2.5; own chart.
- Because shares rose in more than two of three months, simply saying “up” is hard to beat. The best network got it right in 74 of 100 months, many others did worse.
- Invested only when the best network expected rising prices, the money would have grown by 168 per cent instead of 130 per cent with the index. But only 95 of the 648 networks beat the index, and which one is best is known only afterwards.
2020Master’s thesis, LMU MunichExpectations
A comprehensive analysis of the use of deep learning models for forecasting the cross-section of stock returns
Cem Öztürk
Of nine neural networks only two earned more per unit of risk than a simple equally weighted basket of large US shares, before costs.
Nine neural networks: return per unit of risk

All large caps, equal weights0.51
Network with 1 layer0.40
2 layers0.65
3 layers0.43
4 layers0.41
5 layers0.48
6 layers0.44
7 layers0.43
10 layers, variant a0.37
10 layers, variant b0.56

Sharpe ratio of the tenth of shares each network rated best, bought only, January 1976 to December 2018, before costs; US large caps. Source: master’s thesis of Cem Öztürk, LMU Munich, 2020, table 3.7; own chart.
- Tested over 43 years on large US shares, retrained every quarter with 66 indicators; which indicators mattered could be read from the networks: above all book-to-market and price-to-book.
- Only two of the nine beat the simple basket, and precisely these two the selection procedure had not favoured beforehand. Add high trading and deeper falls than the basket.
2020Industry project, UC BerkeleyExpectations
Analysis of stock return predictors using profitability metrics
Alex Bietrix and Paul-Noël Digard
Models that voted each month on the direction of the S&P 500 cut its largest loss from −53 to −22 per cent, with the same return per unit of risk but less return.
Timing the S&P 500: less return, much smaller losses

Return a year
S&P 5007.4%
Timing strategy5.3%
Swings a year (volatility)
S&P 50014.1%
Timing strategy9.7%
Largest loss from a peak
S&P 500−53%
Timing strategy−22%

January 1992 to January 2019, monthly: invested in the S&P 500 or in cash, depending on the vote of the models; long-only variant with a growing data window. Return per unit of risk (Sharpe ratio) 0.58 for both. Source: report of Alex Bietrix and Paul-Noël Digard, UC Berkeley, 2020, table 3; own chart.
- Judged only by profit, not by forecast statistics: each month the models that had earned most over the last two years voted on whether the index would rise.
- No variant beat the index. Of 23 single indicators only one did (nine-month momentum): with a single number the market can hardly be beaten.
2020Industry project, UC BerkeleyEnvironment
Business news topic modeling
Zacharie Bouhnik and William Lambert
The topics of the New York Times explain about 40 per cent of the swings of the US stock market and 16 per cent of its monthly returns, after the fact.
How much of the US stock market the news topics explain

Swings of the S&P 500 (volatility)40%
Monthly return of the S&P 50016%

Share of the month-to-month variation explained by the attention the New York Times gave to five topics in the same month (R²; volatility about 40 per cent). Explained afterwards, not forecast. Source: report of Zacharie Bouhnik and William Lambert, UC Berkeley, 2020, section 4; own chart.
- The basis: 219,162 articles from the business section of the New York Times, 1998 to 2020, sorted by a topic model into 40 topics such as energy, pensions or crisis.
- Attention to the topic “crisis” rose steeply as early as February and March 2007, months before the financial crisis broke. More attention to recession, loan problems and energy went with falling prices.
2020Industry project, UC BerkeleyAllocation
Interpretable models for portfolio management
Jiayin Guo and Yvonne Zhu
Explainable regressions with chosen interactions built better long-short portfolios than decision trees: about half the largest loss and more return per unit of risk.
Long-short portfolio: largest loss from a peak

Regression with interactions−21%
Regression−30%
Random forest−40%
LightGBM−51%

US shares 1980 to 2019, rebalanced monthly: bought the tenth with the best forecasts, sold short the tenth with the worst, equal weights. Return per unit of risk (information ratio): regressions 0.98 and 0.96, random forest 0.61, LightGBM 0.24. Source: report of Yvonne Zhu and Jiayin Guo, UC Berkeley, 2020, figures 3 and 4; own chart.
- The aim was models one can explain: every year the indicators and their interactions were chosen anew, and Shapley values showed why a model picked a share.
- For portfolios that only buy, the decision trees were ahead; for long-short portfolios the simple regressions, and interactions made the falls smaller.
2020Industry project, UC BerkeleyExpectations
LSTM for stock return prediction and portfolio construction
Xuanyi Ji and Shimai Zhang
A neural network (LSTM) ranked US shares by expected return: very accurate in training, much less so in the test.
2021Working paperSelection and review
Modern asset management
Christian Maschner and Martin Schmitz
SSRN
2021Industry project, UC BerkeleyExpectations
Amazon customer reviews and stock returns
Yifei Tong
2021Industry project, UC BerkeleyEnvironment
Ensemble sentiment dictionaries for financial applications
Philip Spalding
2021Industry project, UC BerkeleyExpectations
Forecasting stock market returns with AutoML
Ayush Gupta
2021Industry project, UC BerkeleyExpectations
Forecasting stock market returns with deep learning
Alec Madayan
2021Industry project, UC BerkeleyExpectations
Forecasting stock market returns with ensembles
Kunal Chakraborty
2021Industry project, UC BerkeleyExpectations
Investment strategy with long-history data and feature engineering
Taige Hong
2021Industry project, UC BerkeleyExpectations
Textual analysis of financial statements with ESG language to forecast single-stock returns
Neil Gulati
2022Master’s thesis, TU MunichExpectations
Automatically tuned deep neural networks for forecasting the cross-section of stock returns
Antonio Di Giovanni
2022Master’s thesis, TU MunichExpectations
Empirical asset pricing via machine learning, after trading costs
Dominik Eichhorn
2022Industry project, UC BerkeleyExpectations
Machine learning for equity style rotation strategies
Andrew Lazzeri
2022Industry project, UC BerkeleyExpectations
Stock selection with deep learning
Nathan Sheng
2022Industry project, UC BerkeleyEnvironment
Volatility-managed factor portfolios with machine learning and macroeconomic data
Peter Masforroll
2022Industry project, UC BerkeleyEnvironment
Volatility-managed portfolios with machine learning and macroeconomic data
Divya Singh
2023–2025PhD thesis, LMU MunichStress tests
DCC-GARCH copula models
Dennis Mao
2023Industry project, UC BerkeleyPrivate markets
Cash-flow models for alternative investments
Juliette Ould-Aklouche
2023Industry project, UC BerkeleyGoals
Dynamic portfolio allocation in goal-based wealth management over multiple periods
Elisa Piaraly
2023Industry project, University of CologneMethods and teaching
Machine learning for economists
More than 20 students
2023–2024Industry project, UC BerkeleyStress tests
Modelling the global economy, public and private asset classes for scenario analysis
Angel Chen and Jan-Luca Frick
2023Internship, Frankfurt SchoolAllocation
Multi-period portfolio optimisation
Jan-Luca Frick
2023Industry project, UC BerkeleyAllocation
The Python package riskfolio-lib in depth, with an empirical application
Mélanie Fabiani
2024Master’s thesis, University of ColognePrivate markets
Forecasting net values and cash flows of private assets with machine learning
Janina Dierk
Since 2025PhD thesis, University of Cologne, University of TübingenGoals ongoing
Investor expectations: analysis, forecasting and strategy with artificial intelligence
Thorben Lützen and Felix Klosterköther
Since 2025Industry project, Goethe University Frankfurt, Kiel Institute, University of TübingenExpectations ongoing
Real estate: analysis, forecasting and strategy with artificial intelligence
Tim Brück, Jonas Zdrzalek, Steffen Zetzmann and Felix Klosterköther
Spreading apartments across several German cities would have given, from 1995 to 2025 and at comparable risk, a final value about 27 percentage points higher than a market-weighted portfolio.
GREIX portfolio analysis: spatial diversification pays off (Kiel Institute, in German)
Since 2025PhD thesis, LMU Munich, University of Tübingen, University of ColognePrivate markets ongoing
Private assets: analysis, forecasting and strategy with artificial intelligence
Konstantin Scheurer, Felix Klosterköther and Tom Zimmermann
2026Industry project, UC BerkeleyGoals
Goal-based investor profiling with mental accounts
Stefan Wennemar
2026Industry project, UC BerkeleyEnvironment
Methods for extracting information from financial texts
Jonathan Hartadi
2026Industry project, UC BerkeleySelection and review
Mutual fund selection with machine learning
Bailey Lin
2026Industry project, UC BerkeleyPrivate markets
Subscription lines and forecast bias in the Takahashi-Alexander model
Hrafnhildur Líf Jónsdóttir
Winter 2026/27Lecture, Duale Hochschule Schleswig-HolsteinMethods and teaching ongoing
Capital Markets

How the projects feed into my work with families: Life and financial planning.

Experience

Since 2022
Head of Investment Strategy & Research, member of the Investment CommitteeFINVIA, Frankfurt
- Works out, with his team, the strategic asset allocation across liquid and illiquid asset classes.
- Forecasts returns, risks and correlations, optimises portfolios over several periods, runs scenario analyses and stress tests, and models cash flows and NAVs of alternative investments.
- Leads research collaborations with students and doctoral researchers from international universities.
2018–2022
Executive Partner, part of the founding teamHQ Asset Management, Düsseldorf
- Designed a cloud-based investment platform (R, Python, SQL, Azure, Docker, GitHub) for asset allocation, risk management, equity selection, forecasting, optimisation, backtesting, explainable AI and AutoML.
- Brought artificial intelligence and alternative data into the investment process, led research and portfolio projects across the firm and developed ESG-integrated solutions with MSCI Barra.
- Supervised master’s students on data-driven projects, supported sales and presented research at conferences and universities.
2008–2018
Vice PresidentSal. Oppenheim, Cologne
- Co-led the asset allocation strategy for EUR 7.1 bn of institutional and EUR 1.6 bn of private assets; set up the strategic allocation process and the annual outlook.
- Developed quantitative models for tactical and strategic allocation, including risk simulation, factor allocation and macro indicators.
- Managed five mutual and special funds (EUR 0.5 bn) and contributed to mandate proposals in the investment committee.

Papers

Tree-based conditional portfolio sorts: The relation between past and future stock returns

With Tom Zimmermann. Working paper on SSRN, earlier titled Deep conditional portfolio sorts.

The question it started with: which of the many variables really tell us something about future stock returns? The answer came from artificial intelligence.

By 2013, research had found more than 300 variables said to predict future stock returns. In his 2011 presidential address to the American Finance Association, John Cochrane called this a zoo of factors and asked which of them carry information of their own and which merely repeat others. He suspected that answering this would take different methods.

Double sort, as before

Cuts fixed in advance: 25 groups

Tree, using AI

Cuts chosen from the data: 10 groups

Expected return next monthlowerhigher The same simulated stocks, sorted on a fixed grid on the left and split by a tree on the right. With two variables the grid still works: five times five makes 25 groups. With the 25 variables of the paper it would be nearly 300 quadrillion groups, far more than there are stocks. The tree chooses for itself which variable to cut and where, and needs only 10 groups here. Move the slider to grow it level by level.

How it works, and what came out

The standard tools of the field could not go far. A portfolio sort splits stocks into groups by one variable, a double sort by two; with three or more, each group holds too few stocks. A Fama-MacBeth regression copes with many variables but hardly with their interplay: 50 variables with all pairwise interactions already need 1,275 coefficients, more than there were stocks in some months.

Tom Zimmermann and I answered in 2013 by bringing decision trees and random forests from machine learning, a branch of artificial intelligence, to the stock market. A tree sorts like a multiple sort, but it chooses for itself which variable to split on and where, over as many levels as the data support. The random forest averages 200 such trees, each grown on a random draw of data and variables, so that no chance pattern passes for a law. It was run on Harvard University’s supercomputer.

Applied to the past monthly returns of US stocks since 1963, the method found that almost all of the information sits in the last six months, often nonlinear and in interaction. A trading strategy built on it reached an information ratio, extra return per unit of risk, about twice that of a Fama-MacBeth regression with interactions; trading costs do not explain the gap. It was the first paper to read conditional portfolio sorts as a machine learning method, and in 2015 it received the Best Paper Award of the German Finance Association.

2013
Presented at the Harvard research seminarSlides (PDF)
2014
First working paper and extended presentationPaper (PDF) Slides (PDF)
2015
Best Paper Award at the annual meeting of the German Finance Association
2016
Cited by Cerniglia, Fabozzi and Kolm in The Journal of Portfolio Management 42(5)Article
2017
Cited by Mullainathan and Spiess in the Journal of Economic Perspectives 31(2)Article
2018
Cited by Gu, Kelly and Xiu in Empirical asset pricing via machine learning, later in The Review of Financial Studies 33(5)Paper
2019
Discussed by Stefan Nagel in the Princeton Lectures in Finance, Machine Learning and Asset PricingDiscussions

The relation between stock market risk and return

With Stefan Mittnik. Chapter 2 of the PhD thesis. Presented 2016–2020, among others at the 7th NextGen AI Conference.

PhD thesis (PDF)

Modern asset management

With Christian Maschner and Martin Schmitz.

SSRN

Applications of textual analysis and machine learning in asset pricing

PhD thesis, LMU Munich, 2018. Advisor: Prof. Stefan Mittnik. PLEXUS Advancement Award for Artificial Intelligence in Finance 2020.

Thesis (PDF) Defence slides (PDF)

Full list of papers on SSRN and Google Scholar.

Talks

30 talks since 2013: at central banks and regulators, universities, asset managers and conferences.

Central banks and regulators 5

2015
Federal Reserve BoardWashington D.C. · Tree-based conditional portfolio sorts
2015
Bank of EnglandLondon · Tree-based conditional portfolio sorts
2015
Bank for International Settlements (BIS)Basel · Tree-based conditional portfolio sorts
2015
Norges BankOslo · Tree-based conditional portfolio sorts
2016
Commodity Futures Trading Commission (CFTC)Washington D.C. · Tree-based conditional portfolio sorts

Universities 10

2013
Harvard University, Financial Economics SeminarTree-based conditional portfolio sorts
2014
Harvard University, Financial Economics SeminarTree-based conditional portfolio sorts
2014
University of Zurich, Computational Economics and Finance SeminarTree-based conditional portfolio sorts
2014
LMU Munich, Financial Econometrics SeminarTree-based conditional portfolio sorts
2016
LMU Munich, Financial Econometrics SeminarThe relation between stock market risk and return
2017
LMU Munich, Financial Econometrics SeminarThe relation between stock market risk and return
2018
LMU Munich, Financial Econometrics SeminarThe relation between stock market risk and return
2018
LMU Munich, Colloquium of the Department of StatisticsTree-based conditional portfolio sorts
2021
TU Munich, Advanced Seminar in Finance and AccountingBuilding a modern asset manager with technology and AI
2021
University of Cambridge, Optimal Portfolio Research SeminarBuilding a modern asset manager with technology and AI

Asset managers and banks 4

2014
Sal. OppenheimCologne · Tree-based conditional portfolio sorts
2014
Wellington ManagementLondon · Tree-based conditional portfolio sorts
2015
BlackRockCalifornia · Tree-based conditional portfolio sorts
2017
DWS InvestmentsFrankfurt · Tree-based conditional portfolio sorts

Conferences 11

2015
11th World Congress of the Econometric SocietyMontreal · Tree-based conditional portfolio sorts
2015
22nd Annual Meeting of the German Finance AssociationLeipzig · Tree-based conditional portfolio sorts
2015
9th International Conference on Computational and Financial EconometricsLondon · Tree-based conditional portfolio sorts
2015
6th CEQURA ConferenceMunich · Tree-based conditional portfolio sorts
2015
2nd SOFINE-CEQURA Spring Junior Research WorkshopBavaria · Tree-based conditional portfolio sorts
2016
15th Colloquium on Financial MarketsCologne · Tree-based conditional portfolio sorts
2019
10th CEQURA ConferenceMunich · The relation between stock market risk and return
2019
AI and Data Science in TradingLondon · Building a modern asset manager with technology and AI
2019
The Trading Show EuropeLondon · Building a modern asset manager with technology and AI
2020
The Trading Show EuropeLondon · Building a modern asset manager with technology and AI
2020
7th NextGen AI Conference (textual analysis)Frankfurt · The relation between stock market risk and return

Talks on the papers were partly given by the co-author.

Teaching

Winter 2026/27
Lecturer, Capital MarketsDuale Hochschule Schleswig-Holstein, Business Administration, 5th semester. Valuation, risk and return, interest rates, equities, further markets and portfolio, with an investment committee simulation.
2012
Lecturer, Financial Innovations (MSc)Koblenz University of Applied Sciences

Education

2012–2018
PhD in Statistics (Dr. rer. nat.), LMU MunichAdvisor: Prof. Stefan Mittnik. Dissertation: Applications of Textual Analysis and Machine Learning in Asset Pricing. Completed alongside full-time work.
2004–2008
Diploma in Mathematics and Economics, Koblenz University of Applied Sciences

Summer schools

2016
Advances in Macroeconomic ForecastingGabriel Pérez-Quirós, Madrid
2014
Financial EconometricsEnrique Sentana, Rethymno
2011
Intermediate MacroeconomicsLondon School of Economics
2010
Financial Econometrics and ForecastingFrancis X. Diebold, Geneva

Internships and student jobs

2007–2008
Intern, Quantitative Equity ManagementSal. Oppenheim, Cologne
2006–2007
Intern, Credit Risk Stress TestingDresdner Bank, Frankfurt
2005–2006
Working student, Venture Capital Funds ManagementKfW, Bonn
2005
Intern, Industrial Risk AnalysisAXA, Cologne

Awards

2020
PLEXUS Advancement Award for Artificial Intelligence in FinanceFor the PhD thesis Applications of Textual Analysis and Machine Learning in Asset Pricing; presented in March 2021
2015
Best Paper Award, annual meeting of the German Finance AssociationFor Tree-based conditional portfolio sorts, with Tom Zimmermann

Working together

Students and researchers who would like to write a thesis, join an industry project or work on a joint paper can write to me with a short outline of the idea and the timeline.

Timeline

Financial Innovations (MSc)

Tree-based conditional portfolio sorts

The relation between stock market risk and return

Assessing the risk tolerance of retail investors: risk profiling in theory and practice

Applications of textual analysis and machine learning in asset pricing

Finanzmarktprognose mit Hilfe von Textdaten

Forecasting financial time series with deep learning

A comprehensive analysis of the use of deep learning models for forecasting the cross-section of stock returns

Analysis of stock return predictors using profitability metrics

Business news topic modeling

Interpretable models for portfolio management

LSTM for stock return prediction and portfolio construction

Modern asset management

Amazon customer reviews and stock returns

Ensemble sentiment dictionaries for financial applications

Forecasting stock market returns with AutoML

Forecasting stock market returns with deep learning

Forecasting stock market returns with ensembles

Investment strategy with long-history data and feature engineering

Textual analysis of financial statements with ESG language to forecast single-stock returns

Automatically tuned deep neural networks for forecasting the cross-section of stock returns

Empirical asset pricing via machine learning, after trading costs

Machine learning for equity style rotation strategies

Stock selection with deep learning

Volatility-managed factor portfolios with machine learning and macroeconomic data

Volatility-managed portfolios with machine learning and macroeconomic data

DCC-GARCH copula models

Cash-flow models for alternative investments

Dynamic portfolio allocation in goal-based wealth management over multiple periods

Machine learning for economists

Modelling the global economy, public and private asset classes for scenario analysis

Multi-period portfolio optimisation

The Python package riskfolio-lib in depth, with an empirical application

Forecasting net values and cash flows of private assets with machine learning

Investor expectations: analysis, forecasting and strategy with artificial intelligence

Real estate: analysis, forecasting and strategy with artificial intelligence

Private assets: analysis, forecasting and strategy with artificial intelligence

Goal-based investor profiling with mental accounts

Methods for extracting information from financial texts

Mutual fund selection with machine learning

Subscription lines and forecast bias in the Takahashi-Alexander model

Capital Markets

Experience

Papers

Tree-based conditional portfolio sorts: The relation between past and future stock returns

The relation between stock market risk and return

Modern asset management

Applications of textual analysis and machine learning in asset pricing

Talks

Central banks and regulators 5

Universities 10

Asset managers and banks 4

Conferences 11

Teaching

Education

Summer schools

Internships and student jobs

Awards

Working together