There are some stories that can only take place in Manhattan. A particular kind of crossing over of people and ideas that leads to action of many consequences happens here so often. There is an alternative futures game to be evaluated. What kind of view of credit risk in property loans in early 2000’s would have prevailed if we had used widely machine learning rather than traditional models, and would that have mattered for the outcomes of the 2007–2012 financial crisis? This was an engaging topic for conversation on a slow and long afternoon with old colleagues and friends. The question is possible and current in the first place because at the center of events fifteen years ago, there were deeply hidden dependencies among credit risks of which we knew truly little. The topic is also a favorite among market practitioners, who were active during these momentous times. After much time and effort spent on steering the conversation into calmer and collaborative tones, two scenarios produce two distinct outcomes — one predictable, and one much less clear — a lot more uncertain. So here — we set the stage for the alternative history games to commence. Talk is flowing and so is the sentiment of explorers into alternative futures. In both scenarios we have four types of market players matrixed from two categories. Firstly, players are divided into traditionalists, who employ standard statistical models and machine learners, who leverage artificial intelligence. The second category divides market players into short-term and long-terms investors. Then all-in-all we profile four types of players — traditionalist and long-term investor, traditionalist and short-termist, machine learner and long-termist, and lastly — machine learner and short-termist.
The four types of market players have two parallel verses — starting from two initial states of the market game. In the more stable and predictable game, the one closer to historical reality, all players begin from a state with good command and access to machine learning methodology, implementable for technical pricing of credit default risk. The machine learners are 50% of all players and they have implemented the new algorithms in their market strategies. The other 50% of players — the traditionalists, rely on standard models. However, all players still operate with the same data sets and the same sources available as they were available to us in 2005. In this game machine learners do not explore or do not have access to new data sets and new sources for training their systems. This is a stable and predictable game and here is why. Practitioners using standard models, 50% of our population of players, would come to the same pricing and market-making decisions as were consolidated in 2007. The other cohort, the machine learners, will train their models for pricing and trading with the same data sets as everyone else. They will divide data into two temporal buckets — set [1980–1998] and set [1999–2005], and thus will take advantage of the ability of their systems to do chronological and supervised learning. This is the moment of innovation and the one expected to open a path to an alternative future. The viability of this alternative future can be staked on a solid experiment. The proposition compares Monte-Carlo simulations with parameters derived from a single and full data set [1980–2005] to the machine learning algorithm of random forest with two stages of supervision. Ranked overview would show that the outcomes are not significantly different. The ML approach creates more extreme, unpredictable, and less obvious scenarios, provides wider variability in the whole ensemble, richer in volatility and in quantified dependencies. However, when it comes down to business practitioners ought to converge on centralized and aggregated outcomes. Then it is reasonable to claim that both the traditionalists and the machine learners would have arrived at the same technical profiles of credit risk in property loans’ portfolios, as we did with traditional statistical and simulation methodologies in the early 2000’s. It is already late afternoon and all of us are well pleased by the conclusion, some more than others. Some colleagues sat nearer the fire, and some had all their irons in it. Memories, opportunities, and contingencies bring out emotions and strong claims. Nonetheless there are few sceptics when it comes to the power of machine learning, and all concur that to unlock that might training data needs to be rich and wide.
It is a few of us, a small group gathered after a seminar, and we are looking for a quiet place to sit down somewhere in the East Village. We had strolled for a while, enjoying a warm autumn afternoon in lower Manhattan. It is time to talk about how to evaluate a particular game of alternative futures. It all started on a simple premise at a seminar this week — if machine learning were to be widely adopted by industry practitioners in the wake of the 2007/12 credit crisis, how would have our view of credit default risk for property loans changed and would that have mattered in the forthcoming crisis. To play the game we set the following three rules. Rule number 1 — although machine learning methodology is widely available and implementable, half of market players have adopted it for their production systems and the other half continue to use the traditional models available and dominant at that time, (i.e., Monte Carlo simulation of various Gaussian methodologies). This is the first piece for setting up the alternative future game. Rule number 2 is next. The game is played among four groups of practitioners matrixed by two categories — firstly, players are divided into traditionalists, who employ standard statistical models and machine learners, who leverage machine learning in artificial intelligence. The second category divides market players into short-term and long-term investors. Then all-in-all we profile four types of players — traditionalist and long-term investor, traditionalist and short-termist, machine learner and long-termist, and lastly — machine learner and short-termist.
A game set on these two rules was deemed stable and predictable. We played it and managed to sort out the various outcomes. The group concluded that the ML approach creates more extreme, unpredictable, and less obvious scenarios, provides wider variability in the whole ensemble, richer in volatility and in quantified dependencies. However, when it comes down to business practitioners will converge on averaged and aggregated outcomes. We claimed that in the end both the traditionalists and the machine learners would have arrived at the same technical profiles of credit risk in property loans’ portfolios, as we did with traditional statistical and simulation methodologies in the early 2000’s.
This clears the stage for the more complex and less predictable game. All the premises of the first game are present. In addition, there is rule number 3 — the market and practitioners have access to new data sets and sources for estimating and training their credit default risk models. In the first game, as in the reality of the early 2000’s, credit default risk was derived from data of historic defaults in property loans. In this new scenario historic defaults of high yield corporate bonds and municipal bonds are available as new and additional data sources. Both groups of players, the traditionalists and the machine learners have access to this new data. The traditionalists will find it hard to impossible to incorporate in their standard models for credit default, parameters estimated from risks, such as corporate and high-yield bonds. These risks are external to their standard models. Their models are purposefully designed to take parameters only from internal (to the modeling process) risk factors and cannot accommodate variables gleaned or estimated from ‘external’ risks.
On the other hand, machine learning systems are perfectly capable of training themselves with and without human supervision from a targeted data set. Market practitioners using them will embrace the opportunity to leverage this might. Data on default rates of high-yield bonds in 2000–2002 can then be put to clever use to train the ML system to uncover a threshold with a signal for a system-wide spike, in the likelihood of credit defaults. Such a signal on a critical threshold for a systemic cascading effect is not detectable in standard data of default rates in property loans. However, the tremors will be picked up from the high-yield bonds’ data. This is the moment of market innovation, and this is the trigger that gives advantage to the ML practitioners both for their own benefit and for the benefit of systemic stability in the whole market.
What happens next in this alternative future, we simulate diligently with a genetic algorithm. To begin with 33% of credit risk transfer transactions will not occur in this imaginary market, because of a large discrepancy between the views of traditionalists and machine learners. Hence, the overall market itself will be 33% smaller, thus significantly decreasing exposure to default for all market players and the financial system. Another 33% of credit risk transfer is priced by practitioners at a lower premium than it was done in 2007, reflecting higher likelihoods of credit defaults, and thus decreasing market risk for all parties. ML practitioners armed with the ability to estimate dependencies and thresholds for cascading systemic shock from relevant new data sources reduce both exposure and market risk for two-thirds of the active market.
Schema: Machine learning guides market transactions in credit default risk
Despite the employment of new methodologies and new data sources, one third of the market for risk transfer is overpriced in transactions among the players. We need to allow for all parties, even those using AI systems with sourced learning, that they will never be fully protected from the pitfalls of inaccurate technical pricing. Thus, the market price for one third of all transactions remains volatile, precisely because it is not at technical equilibrium. As a result, in a second round of trading, the price of credit default risk is discounted downwards until it finds its new stable state. With this our experimental market game is complete. Two events have occurred in our game of alternative futures that did not occur in the historical reality of 2007–2012. The overall market for credit risk transfer was smaller because players were less certain of the technical price of risk. This is a good thing and it reduced exposure risk. Many transactions were priced lower than the historical average, because ML systems accounted for market experiences with the high-yeild bonds market. This reduced further market risk. And one event repeated itself to the historical reality. With powerful algorithms, mistakes will still occur and as a results one third of risk trasnfer was still not priced well. As in 2008 so in the game the volatility of these transactions caused a market shock. The shock is much smaller and therefore much easier to absorb by the credit risk market.
Would this still lead to a financial crisis and a economic recession is a subject for anaother round of experimentation and well-supervised learning. The machines have showed us that they can learn. The craft of what and how to learn and the art of putting the product to good use is still up to the humans.