From Goldman Sachs to Europe's Largest Digital Finance Network

Nihar:

Welcome back to the Amsterdam Investment Club Podcast. I'm your host, Nihar Mahesh Jani, and today we have a truly fascinating guest - someone who has walked the floors of Goldman Sachs, built trading algorithms at Man AHL, and now leads one of Europe's largest research networks in digital finance. Professor Joerg Osterrieder, welcome to the show.

Joerg:

Thank you, Nihar. It's a pleasure to be here. I've been following the Amsterdam Investment Club for some time, and I really appreciate what you're doing to connect students with professionals in our field.

Nihar:

Let's start at the beginning. You have a background in mathematics from ETH Zurich, Syracuse University, and business economics from Ulm. That's quite an international education. How did that shape your approach to finance?

Joerg:

You know, I think the international exposure was crucial. At ETH, I got rigorous mathematical training - the Swiss approach is very thorough, very precise. At Syracuse, I was exposed to the American way of doing applied mathematics, which tends to be more pragmatic, more focused on solving real problems. And Ulm gave me the business context - understanding not just the mathematics but why it matters for real organizations. That combination served me well when I moved into industry because finance requires all three: mathematical rigor, practical problem-solving, and business understanding.

Nihar:

Your first role in investment banking was at Merrill Lynch in 2007. That's a significant time to enter the industry - right before the financial crisis. What was that experience like?

Joerg:

It was intense, to say the least. I joined as an Associate in Global Markets in London, working on algorithmic execution strategies. We were building systems to optimize how large orders get executed in the market - minimizing market impact, reducing transaction costs. The work itself was fascinating: taking theoretical ideas about optimal execution and implementing them in production systems that handled real money. But of course, 2008 changed everything. Watching the crisis unfold from inside a major investment bank was... educational, in ways I couldn't have anticipated. You learn very quickly what systemic risk actually means when you see it happening around you.

Nihar:

In 2009, you moved to Goldman Sachs as a Vice President in Global Markets. What drew you there, and what did you work on?

Joerg:

Goldman had a reputation for having the best technology in the industry, and that was largely true. The infrastructure they had built for algorithmic trading was years ahead of most competitors. I was drawn by the opportunity to work on more sophisticated systems with some of the brightest people in the field. My role focused on developing and implementing algorithmic execution strategies - VWAP, TWAP, implementation shortfall algorithms. We were building on the Almgren-Chriss framework for optimal execution, minimizing market impact while achieving target prices. The math is elegant: you're balancing the risk of price drift against the cost of trading too aggressively and moving the market against yourself.

Nihar:

Can you explain VWAP and TWAP for listeners who might not be familiar?

Joerg:

VWAP is Volume-Weighted Average Price - you're trying to execute at the average price weighted by volume throughout the day. If 30% of volume trades in the first hour, you execute 30% of your order then. TWAP is Time-Weighted Average Price - simpler, you just spread your order evenly across time. But the real sophistication comes from market impact models. When you trade 100,000 shares, you move the price. The Almgren-Chriss model gives you a framework for thinking about this trade-off: execute fast and you create permanent and temporary price impact; execute slow and you're exposed to price drift. Finding that optimal trajectory - that's where the math meets real money.

Nihar:

How did you work with the traders on the desk?

Joerg:

That's where the job gets really interesting - and it's something most academic courses don't prepare you for. A quant's job isn't just writing algorithms - it's explaining to traders why the model behaves a certain way. When our VWAP algorithm suggested holding back during high volatility, I had to explain the market impact model in terms they could act on. 'The algo sees the spread widening and participation rate spiking - if we push now, we'll move the market 20 basis points against us. Let's wait for the liquidity to come back.' That translation skill matters as much as the math.

Nihar:

How do you build trust with traders who might be skeptical of quantitative approaches?

Joerg:

You don't earn credibility on a trading desk by being the smartest person in the room. You earn it by listening, by understanding their P&L pressures, and by delivering code that works when it matters. It took months before traders would trust my signals. I remember one trader who initially dismissed our market impact estimates - until we tracked a month of his manual executions against our model's predictions. The data showed he was consistently underestimating slippage on large orders. After that, he became one of our strongest advocates. Trust is built through results, not presentations.

Nihar:

What was it like working with both humans and algorithms in real-time?

Joerg:

The best moments were when we worked in real-time together - the algo processing market data at microsecond latency, the trader adding context about a news event or a large order they spotted, both adjusting on the fly. I remember a day when the algo flagged unusual activity in the order book - systematic selling pressure that didn't match normal patterns. The trader recognized it as likely program trading from a specific source. Together, we adjusted our execution trajectory to avoid being run over. Neither human nor machine alone would have made the same decision. That synergy - human intuition plus algorithmic precision - that's when the magic happens.

Nihar:

What was the difference between research work and production systems?

Joerg:

That's a crucial distinction that many people don't fully appreciate. In research, you can make simplifying assumptions, you can ignore edge cases, you can work with clean data. Production is completely different. Your VWAP algo needs to handle market halts, data feed failures, extreme volatility - all while maintaining microsecond-level latency. A strategy that looks brilliant in backtesting might fail completely in production because it can't handle the messiness of real markets. I've seen many promising ideas die in the transition from research to production. Understanding both sides - the theoretical elegance of optimal execution theory and the practical engineering of latency-sensitive systems - is what separates good quants from great ones.

Nihar:

After Goldman, you moved to Credit Suisse in 2012, but in a different role - regulatory projects. That seems like quite a shift.

Joerg:

It was deliberate. After 2008, the regulatory landscape was transforming completely. Basel III, Dodd-Frank, MiFID II on the horizon - understanding these frameworks was becoming essential for anyone in finance. I wanted to see the industry from that perspective, to understand how regulations are implemented, what compliance actually means in practice. It was only a brief stint, but it gave me insight into a dimension of finance that pure quants often overlook. When you're designing trading systems, you need to understand the regulatory constraints you're operating within. That knowledge has proven valuable throughout my career.

Nihar:

And then you moved to Man AHL, one of the most respected systematic trading firms in the world. How was that different from investment banking?

Joerg:

Man AHL was a different beast entirely. At banks, algorithmic trading is one activity among many. At AHL, systematic investing is the core business - it's all they do. The depth of research, the sophistication of their systems, the long-term thinking - it was impressive. I worked on multi-asset trading strategies, particularly around risk management and downside protection. We were building products like the Man AHL Target Risk fund, which uses dynamic allocation to manage portfolio volatility. It's elegant work: combining trend following, risk parity, and tail risk hedging into coherent investment products.

Nihar:

What's the philosophy behind systematic investing at a firm like AHL?

Joerg:

The fundamental belief is that markets have persistent patterns that can be exploited systematically - but those patterns are often subtle, noisy, and require disciplined execution to capture. The human tendency is to override systems when they're underperforming, to second-guess signals, to let emotions interfere. AHL's approach is to build robust systems and then trust them, even through difficult periods. That discipline is harder than it sounds. I've seen many traders who can't resist intervening, and often that intervention destroys value rather than creating it.

Nihar:

You had an enviable industry career - Goldman, Man AHL, working on cutting-edge systems. Why leave for academia?

Joerg:

It's a question I get often, and I understand why it surprises people. The honest answer is that I wanted to work on longer-term questions. In industry, you're focused on the next quarter, the next product launch, the next regulatory deadline. There's rarely time to step back and ask fundamental questions: Why do markets behave this way? What are the limitations of current approaches? How should we be thinking about AI in finance? Academia gives you that space. And teaching - helping the next generation develop their skills and thinking - that's deeply rewarding in a way that's hard to replicate in industry.

Nihar:

You currently hold positions at both the University of Twente and Bern Business School. How do you balance those?

Joerg:

It requires careful planning, but the two roles complement each other well. At Twente, I'm an Associate Professor of Finance and AI, so the focus is on research at the intersection of machine learning and financial applications. At Bern, I'm Professor of Finance and Sustainable Finance, which has a stronger emphasis on ESG integration and responsible investment. Both are part-time positions, which allows me to maintain industry connections as well - I'm an Advisor to ING's Global Data Analytics Team, for example. The variety keeps things interesting and ensures my research stays relevant to practice.

Nihar:

The ING advisory role - what does that involve?

Joerg:

I help them think about AI applications in banking: credit risk modeling, fraud detection, customer analytics. Banks are sitting on enormous amounts of data, and there's huge potential to use machine learning to improve outcomes - both for the bank and for customers. But there are also significant challenges around explainability, fairness, and regulatory compliance. My role is to help navigate those challenges, to bring academic rigor to practical problems.

Nihar:

Let's talk about the MSCA Digital Finance program. For listeners who aren't familiar, can you explain what it is?

Joerg:

Certainly. MSCA stands for Marie Sklodowska-Curie Actions - it's the European Union's flagship program for doctoral training. Our project, called Digital Finance, is an Industrial Doctoral Network funded with approximately EUR 3.8 million. We have 17 fully-funded PhD positions across 10 universities and 11 industry partners. The program runs from January 2024 to December 2027. I coordinate the network from the University of Twente.

Nihar:

Who are the partners involved?

Joerg:

On the academic side, we have University of Twente as coordinator, WU Vienna, Poznan University of Technology, Babes-Bolyai in Romania, Kaunas University of Technology in Lithuania, RPTU in Germany, University of Pavia and University of Naples Federico II in Italy, American University of Sharjah, and Bern University of Applied Sciences. On the industry side, we have some remarkable partners: Fraunhofer Institute, Deutsche Bank, Deutsche Borse, Raiffeisen Bank, Swedbank, the Bank for International Settlements, EIT Digital, Royalton Partners, Quoniam Asset Management, Cardo AI, and Athena Research Center.

Nihar:

That's an impressive consortium. What are the research objectives?

Joerg:

We're working to establish Digital Finance as a standalone academic discipline. The research is organized around five interconnected objectives. First, understanding data quality in finance - how do we ensure the data we use for AI models is reliable and representative? Second, advancing AI and machine learning applications in finance. Third, and this is crucial, explainability of AI systems - the XAI agenda. In finance, you can't just say 'the model says so' - you need to explain why. Fourth, blockchain and distributed ledger applications in finance. And fifth, sustainable finance and ESG integration. Each PhD candidate works at the intersection of these themes.

Nihar:

For someone considering applying to such a program, what do they gain?

Joerg:

Several things. First, full funding - salary, research budget, travel allowance - for the entire PhD duration. Second, secondments at industry partners, so you get real-world experience alongside academic training. Third, access to a network of researchers across Europe and beyond. Fourth, training events, summer schools, workshops that develop both technical and soft skills. And fifth, a degree from a respected European university with strong industry connections. It's a pathway that opens doors in both academia and industry.

Nihar:

You also chair a COST Action - CA19130 - on FinTech and AI in Finance. Tell us about that.

Joerg:

COST Actions are European research networks that bring together scientists across borders. I've been chairing this Action since 2020, and it runs until May 2025. When we started, we had perhaps 50 academics interested. Today we have over 400 from 49 countries. The future of finance research isn't built in isolation - it's built through networks. That growth taught me that Europe has a unique opportunity to lead in digital finance, but only if we collaborate across borders.

Nihar:

That's remarkable growth - from 50 to 400. How did that happen?

Joerg:

It wasn't obvious at the start that it would work. There was initial skepticism about cross-border collaboration - different academic cultures, different research traditions, language barriers. The first breakthrough came when we organized a joint publication with researchers from five countries. Everyone brought their expertise: someone had the data, someone had the methodology, someone understood the regulatory context. The paper was better than any of us could have written alone. After that, word spread. Each member brought their network. A researcher from Lithuania would mention the Action to a colleague in Portugal. The multiplier effect was extraordinary.

Nihar:

What kinds of activities has the network organized?

Joerg:

We've had training schools where PhD students from across Europe spend a week learning together - that builds relationships that last entire careers. Short-term scientific missions allow researchers to visit labs in other countries. We've organized conferences and workshops in locations from Lisbon to Vilnius. And crucially, the Brussels workshops - bringing academics face-to-face with policymakers. When the European Commission was drafting technical standards for the AI Act, they wanted to hear from researchers who actually build these systems. Our network could provide 50 experts on a specific technical question within a week.

Nihar:

What's the tangible impact you've seen?

Joerg:

Joint publications, absolutely - hundreds of co-authored papers across the network. PhD co-supervisions across borders - a student in Romania supervised jointly by faculty in the Netherlands and Italy. Policy consultations where our members provided input on MiCA, the AI Act, and digital euro discussions. But perhaps most importantly, we've built a community. When a researcher in Greece has a question about XAI in finance, they know exactly who to email in Germany or Sweden who's worked on that problem. That informal knowledge network might be the most valuable output of all.

Nihar:

What's the focus of the research?

Joerg:

We organize around three transparency angles. First, transparency in FinTech itself - how do we ensure that financial technology serves consumers fairly? Second, the black box versus transparent models debate - when is it acceptable to use opaque AI systems, and when do we need interpretable alternatives? Third, transparency in investment product performance - helping investors understand what they're actually getting. These questions cut across disciplines: computer science, finance, law, economics, even philosophy. That's what makes the network so valuable - bringing together diverse perspectives.

Nihar:

You mentioned engagement with European institutions. Can you elaborate?

Joerg:

Yes, this has been one of the highlights. We've organized workshops in Brussels with participation from the European Commission and European Parliament. We've been involved in discussions around the AI Act - which is now law - and MiCA, the Markets in Crypto-Assets regulation. When policymakers are drafting these frameworks, they want academic input on technical questions. Our network provides that expertise. It's gratifying to see research translate into policy that affects millions of people.

Nihar:

What's your view on where European AI regulation is heading?

Joerg:

Europe is taking a more precautionary approach than the US or China, which I think is appropriate for high-stakes domains like finance. The AI Act categorizes AI systems by risk level, with strict requirements for high-risk applications. Finance falls squarely in that category. Explainability isn't just a nice-to-have in financial AI - it's becoming a regulatory requirement. If you can't explain why your model made a decision, you can't use it in regulated markets. That's pushing the field in interesting directions - toward inherently interpretable models rather than post-hoc explanations.

Nihar:

Let's get into some of your recent research. What are you working on currently?

Joerg:

I have several active streams. One is on credit risk and machine learning - specifically, how can we use ML techniques to extend credit to people who don't have traditional credit histories? This is published in Information Processing and Management. Another stream looks at commodity price modeling using machine learning, published in Decisions in Economics and Finance. We're also working on predicting retail customer distress - identifying customers who might struggle with loan repayments before they actually default - that's in the Journal of Retailing and Consumer Services.

Nihar:

What specific ML methods are you using for credit risk?

Joerg:

We've found that gradient boosting methods - particularly XGBoost - consistently outperform traditional logistic regression for credit scoring. But performance alone isn't enough in finance. You need explainability. We use SHAP values - that's SHapley Additive exPlanations - to decompose every prediction into feature contributions. The regulator walks in, asks why we denied a loan - we show them exactly which features drove that decision. Income-to-debt ratio contributed negative 0.3, payment history on utilities contributed positive 0.2, and so on. That's not optional anymore under GDPR's right to explanation.

Nihar:

Can you explain the difference between SHAP and LIME for our listeners?

Joerg:

Both are explainability tools, but they work differently. LIME - Local Interpretable Model-agnostic Explanations - approximates your complex model with a simpler linear model around each prediction point. It's fast but can be unstable. SHAP is based on game theory - Shapley values from cooperative game theory - and gives you theoretically grounded feature attributions that sum to the prediction. SHAP is more computationally expensive but more consistent. For regulatory purposes, we prefer SHAP because the attributions are additive and satisfy certain fairness axioms. When you're explaining decisions to regulators, you want mathematical rigor behind your explanations.

Nihar:

How do you validate these models? I imagine standard cross-validation might be problematic with time series data.

Joerg:

Exactly right. Standard k-fold cross-validation would leak future information into your training set - you'd be training on 2024 data to predict 2023 defaults, which is cheating. We use walk-forward validation: train on January to December 2022, validate on January to March 2023; then train on January 2022 to March 2023, validate on April to June 2023; and so on. This mimics how the model would actually be deployed. You're always predicting the future using only the past. It's more demanding - your out-of-sample performance will typically be worse than naive cross-validation - but it's honest.

Nihar:

What about feature engineering for credit risk?

Joerg:

Feature engineering is where domain knowledge meets data science. For credit risk, we create time-series aggregations: average transaction amount over 30/60/90 days, trend in account balance, volatility of income deposits. Behavioral features matter: how often does someone check their balance? Do they pay bills early or at the last minute? The machine learning algorithm finds patterns, but the features you give it determine what patterns it can find. Raw transaction data won't tell you much. But 'average weekend spending as percentage of monthly income' - that's predictive. Good feature engineering requires understanding both the data and the business problem.

Nihar:

What about high-frequency trading? I saw you have recent work on reaction times to economic news.

Joerg:

Yes, that's a working paper on SSRN. We're looking at how quickly different types of market participants react to macroeconomic news releases. The speed of reaction has decreased dramatically over time - we're talking milliseconds now. What's interesting is how different actors have different reaction profiles. Some are clearly algorithmic, reacting almost instantly. Others are slower, presumably human or less sophisticated systems. Understanding this heterogeneity helps us think about market microstructure and fairness.

Nihar:

And the practical applications of this research?

Joerg:

For practitioners, understanding reaction times helps with execution strategy - how to get orders filled without being adversely selected by faster traders. For regulators, it raises questions about market design - should there be speed bumps? How do we ensure markets are fair? For academics, it's a window into how information gets incorporated into prices. These questions matter because efficient, fair markets benefit everyone.

Nihar:

For students listening who want to break into quantitative finance, what skills should they focus on?

Joerg:

Programming is essential - Python has become the lingua franca of quantitative finance, but also R for statistical analysis and SQL for working with data. Mathematical foundations matter: linear algebra, probability, statistics, optimization. These aren't just boxes to check - you need real comfort with these tools. But technical skills aren't enough. Communication is crucial. Can you explain your work to traders, to risk managers, to regulators? Can you write clearly? Can you present persuasively? The best quants I've worked with combine technical depth with the ability to translate that into business impact.

Nihar:

Is a PhD necessary to get into this field?

Joerg:

Not necessarily. A PhD is one pathway, and it's valuable for certain roles, but it's not the only route. Many successful quants have master's degrees or even bachelor's degrees with strong technical skills. Industry certifications like CFA or FRM can help. Contributing to open source projects demonstrates your abilities. Research assistantships, even as an undergraduate, can provide valuable experience. The key is demonstrating competence - how you demonstrate it matters less than that you do.

Nihar:

What about students who face financial barriers to pursuing advanced education?

Joerg:

This is something I care deeply about. The biggest barrier to entering quantitative finance isn't talent - it's access. Programs like MSCA exist precisely to break down those barriers, to say: 'If you have the skills and the drive, we'll fund your journey.' Our COST Action has connected over 400 academics, many from countries where research funding is limited. Online resources have democratized learning in ways that weren't possible when I started. Platforms like Coursera, edX, even YouTube have high-quality content available for free. The playing field isn't level yet, but it's more level than it's ever been.

Nihar:

Any specific advice for students from non-traditional backgrounds?

Joerg:

Build a portfolio. Write code, put it on GitHub. Do projects, write about them. Network actively - reach out to people whose work you admire, attend conferences, join online communities. Apply for positions even if you don't meet every criterion - many job postings describe ideal candidates, not minimum requirements. And be persistent. Rejection is part of the process. Every successful person I know has faced setbacks. What matters is how you respond to them.

Nihar:

As we wrap up, where do you see digital finance heading in the next decade?

Joerg:

I see several trends converging. AI will become more embedded in financial services, but with greater emphasis on explainability and fairness. Blockchain technology will mature - we'll move beyond the hype to practical applications in settlement, identity, and compliance. Sustainable finance will become mainstream, not a niche. And regulatory frameworks will continue to evolve as policymakers catch up with technological change. The role of researchers is to anticipate these changes, to provide evidence for good policy, and to train the next generation of professionals who will navigate this evolving landscape.

Nihar:

For listeners who want to learn more about your work or get in touch, where should they go?

Joerg:

My personal website, joergosterrieder.com, has information about my research and teaching. For the MSCA Digital Finance program, visit digital-finance-msca.com - applications for future cohorts will be announced there. The COST Action FinAI has a wiki at wiki.fin-ai.eu with resources, publications, and information about joining. I'm also active on LinkedIn if people want to connect directly. And of course, the University of Twente and Bern Business School websites have information about our programs.

Nihar:

Professor Osterrieder, this has been an incredibly informative conversation. Thank you for sharing your journey and insights with our audience.

Joerg:

Thank you, Nihar. It's been a pleasure. To your listeners: the field of digital finance is wide open. Whether you come from math, computer science, economics, or somewhere else entirely, there's a place for you. Stay curious, keep learning, and don't let barriers - real or perceived - stop you from pursuing what interests you.

Nihar:

And that's a wrap for today's episode of the Amsterdam Investment Club Podcast. If you enjoyed this conversation, please subscribe, leave a review, and share with others who might benefit. Until next time, I'm Nihar Mahesh Jani. Keep investing in knowledge.