# Data Assets & Alpha Group: Demystifying AI in the Financial Domain (Part 1)

- **Show:** Making Sense (J.P. Morgan)
- **Published:** 13 Dec 2023
- **Source:** https://marketmatters.podbean.com/e/market-matters-data-assets-alpha-group-demystifying-ai-in-the-financial-domain-part-1/
- **Transcript:** https://podcasts.opennash.com/shows/jp-morgan-making-sense/data-assets-alpha-group-demystifying-ai-in-the-financial-domain-part-1-1545c9/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Machine learning in investing is best understood as a ladder of models, from simple to complex. Peng Cheng and Emma Wu argue that linear regression remains the quant workhorse. Nonlinear models such as random forests and support vector regression are worth adopting when there is an economic reason for nonlinear behaviour, or when linear models fail out of sample. Each model has costs in interpretability and computing time.

The central tension is between power and practicality. More flexible models can capture real market patterns, such as the way equity returns relate to volatility. But they can be slow, hard to explain, or poorly suited to continuous financial data. The guests also say finance should relax its old habit of minimizing the number of variables, because modern computing power makes extra features cheap to keep.

## Key insights

1. **Linear regression is still the quant workhorse** Peng Cheng says it is easy to understand, extend and implement, so it should be the first tool tried. Nonlinear models should be used only with a strong economic rationale or when linear models fail out of sample.
2. **Machine learning tolerates more variables than statistics** Peng says statistics prizes parsimony, meaning few parameters, partly because computing power was limited. He says 30 variables on 1,000 observations is fine, because extra features often help out-of-sample forecasts.
3. **Ensemble methods differ in how models are combined** Emma Wu uses a portfolio manager and analysts analogy. Bagging has analysts work on different data subsets and average the results. Boosting has each analyst focus on the previous one's errors. Stacking has a lead analyst combine models built by different methods.
4. **Random forest beat linear models in an FX hedge** Emma's team started with about 30 signals and narrowed them to four using recursive feature selection. Random forest outperformed the other models they tested, linear models included, in their tail-hedge FX options model.
5. **Support vector regression captures curves but scales badly** Peng says it handles continuous nonlinear relationships, such as returns versus volatility, and resists overfitting. Its training time grows steeply as the dataset grows, which limits its use on large problems.

## Questions this episode answers

### What is the difference between AI and machine learning in investing?

Peng Cheng says machine learning is any quantitative model that produces investment signals, and it learns its own parameters from data rather than following fixed rules. AI is the broader field, which also handles text, images and voice, as ChatGPT does.

### Is linear regression really machine learning?

Yes, according to Emma Wu, who calls it the fundamental machine learning algorithm. Peng Cheng adds that the difference from traditional statistics is more cultural than technical, since one field grew out of computer science and the other out of statistics.

### What does overfitting mean, and how do you prevent it?

Overfitting is when a model predicts its training data well but fails on new data, which Emma compares to memorizing answers without understanding. One fix is regularization, which adds penalties to the model's scoring function so it does not fit noise.

### How do random forests and boosting differ?

A random forest is a collection of decision trees, each trained on a different subset of the data, whose predictions are averaged. This is called bagging. Boosting builds models in sequence, and each new model concentrates on the mistakes of the one before it.

### When should an investor choose a nonlinear model over a linear one?

Peng says to move to nonlinear models only if there is a strong economic reason to expect nonlinear relationships, or if linear models do not perform well out of sample. He also warns that tree-based models handle continuous financial relationships poorly and can be costly to run.

## Facts

- **Peng Cheng** [0:02:10]: "But machine learning will actually learn these parameters by itself based on the data and find out what their optimal values are."
- **Peng Cheng** [0:03:13]: "machine learning can be thought of as a subset of AI"
- **Eloise Goulder** [0:04:17]: "the more complex the model, and indeed, if the model is non-linear, then it is much more likely to lead to a more accurate relationship"
- **Emma Wu** [0:05:04]: "the model aims to find the best feed line that minimizes the sum of square differences between the actual and predicted values."
- **Emma Wu** [0:05:34]: "there is a situation that the model gives accurate predictions for training data, but not for the new data."
- **Emma Wu** [0:05:34]: "regularization is like you add penalties to the loss functions"
- **Emma Wu** [0:08:47]: "ElasticNet combines the penalties of both Lussell and Rage so that it can be useful when the data set is large and when there are large numbers of features"
- **Peng Cheng** [0:06:42]: "historically, statistics have always valued parsimony, which means to have as few parameters, as few coefficients as possible in the model"
- **Peng Cheng** [0:07:58]: "it's totally okay if we have a thousand observations and 30 explanatory variables."
- **Peng Cheng** [0:10:44]: "if I want to track a stock with a portfolio of stocks, sometimes I cannot have 500 stocks tracking this one stock."
- **Emma Wu** [0:12:28]: "random forest is essentially a collection of decision tree"
- **Emma Wu** [0:15:07]: "the next model or the next member, when going into the model, consistently emphasized the error from the previous member or the model."
- **Emma Wu** [0:15:44]: "you assign a lead analyst that you trust most to take the predictions of each models as the input and make a meta model to make the final predictions."
- **Emma Wu** [0:17:13]: "So last year, we developed a systematic tailhatch model to apply FX options selectively in risk of market and we call it a tailhatch late cycle currency model."
- **Emma Wu** [0:17:13]: "initially we start with a list of about 30 signals across different categories such as value, momentum, swap rates, and volve factors."
- **Emma Wu** [0:18:25]: "we find that linear underperformed nonlinear models with a relatively higher tracking error and poor explanatory power."
- **Peng Cheng** [0:19:45]: "if you think about trees, they are pretty good at capturing nonlinear relationships, but only the discrete ones, right?"
- **Emma Wu** [0:21:45]: "the support vector term stands for the data points that closest to the decision boundary."
- **Emma Wu** [0:22:18]: "Hyperparameters in machine learning are like settings on a recipe that you can adjust before cooking."
- **Peng Cheng** [0:23:27]: "It's been developed in the 1990s, so packages are easy to find, they're easy to implement."
- **Peng Cheng** [0:23:55]: "in periods of low volatility, equity returns are probably good, right? But as volatility picks up, then equity returns declines."
- **Peng Cheng** [0:24:26]: "by using SVM, It actually produces much better out-of-sample performance than a number of other machine learning models."
- **Peng Cheng** [0:24:26]: "it's very computationally heavy, and it doesn't really scale well to large problems."

## Transcript

**[0:00:02] Speaker 4:** Welcome to Market Matters, our markets podcast on making sense, the hub for JPMorgan corporate and investment bank podcasts. In each episode of Market Matters, we discuss the latest news

**[0:00:12] Eloise Goulder:** and trends shaping markets today. Hi, I'm Eloise Goulder, head of the Data Assets and Alpha Group here at JPMorgan. And today I'm really pleased to be sitting here in person in New York with Peng Cheng, who is head of big data and AI strategies in our global research team, and with Emma Wu, who is a member of Peng's team and has done a lot of work on AI strategies. So Peng, Emma, thank you so much for joining me here today.

**[0:00:43] Peng Cheng:** Hi, Eloise. It's great to be back.

**[0:00:45] Eloise Goulder:** Hi, Eloise. Hi, Peng. Thanks for having me. So Peng, Emma, I'm really looking forward to this discussion with you on demystifying AI in the financial domain. I think we can all agree that this has been a hugely popular topic this year across the media, across corporates, across big tech, and also in the academic sphere. And there's a huge amount of information and content out there already. But what we really wanted to achieve today was to demystify some of the popular AI models and keywords, and importantly, to put them into the context of our world, which is the investing space.

**[0:01:24] Eloise Goulder:** So Peng, can we start very much with the basics? And in the spirit of demystifying, can you start by explaining what AI and ML are and what the

**[0:01:35] Peng Cheng:** difference between those two terms are? Yeah, sure. So as you can imagine, this is a very broad topic, but we'll just focus on our domain, which is quant investing. So first of all, machine learning is any quantitative models that produces investment recommendations and investment signals, right? That's how I would define it. How does it differ from, let's say, other quantitative models? So, for example, you can have a rule-based model. An example would be a momentum strategy, which may say, okay, I will observe the 10-day, 30-day moving average, and if one crosses over the other, I will buy.

**[0:02:10] Peng Cheng:** Otherwise, I will sell. So, this is a rule-based strategy where I specify all the parameters. But machine learning will actually learn these parameters by itself based on the data and find out what their optimal values are. So for example, I don't necessarily have to look at 10-day, 30-day moving average, maybe nine and a half day, 28 days or better. And I don't necessarily have to have only one set of moving average crossovers. I can have five, 10 sets, and those are all determined by the model itself based on data.

**[0:02:44] Peng Cheng:** And how does this differ from the traditional econometric analysis? Well, actually, from my perspective, I don't think there's such a huge difference. It's probably more philosophical than technical. You know, one originates from computer science and the other originates from statistics. So a lot of the differences mainly come from things like terminology. So we can go into the details later when describing these specific models. And lastly, you also mentioned AI. So what's the difference between AI and machine learning?

**[0:03:13] Peng Cheng:** I think, generally speaking, machine learning can be thought of as a subset of AI. So when we talk about AI, it's usually perceived to include not only numerical inputs and outputs, but also things like text, images, voices, right? When we think about ChatGPT, it produces words instead of numbers. So that's, generally speaking, what we think of AI as it's more of a superset of machine learning.

**[0:03:36] Eloise Goulder:** That's really helpful. Thank you, Peng. So can we turn to the meat of this discussion then, which is the models? Peng, can you start by listing the main major models out there?

**[0:03:47] Peng Cheng:** Yeah, sure. So there are probably an infinite number of models out there, but we'll just pick a few that are representative and go from the most simple to the most complex. So we'll start with variations of linear regression, and then trees and random forests, support vector machines, and support vector regressions. So these will be considered classical machine learning models. And then we can move on to the deep learning models, namely some common-used architectures of neural networks, and finally, on to large language models.

**[0:04:17] Eloise Goulder:** So that sounds like a great plan, Peng, and really interesting that you're ordering these from the most simple to the most complex, because I remember in our last conversation together on this podcast series, it was back in mid-September, you were arguing there that generally in your research, you found that the more complex the model, and indeed, if the model is non-linear, then it is much more likely to lead to a more accurate relationship between input variable and output and therefore in general is much more powerful in forecasting returns and creating alpha.

**[0:04:54] Eloise Goulder:** So the first broad set of models that you just described Peng were linear regression so let's start with that one and Emma can we turn to you now could you start by defining the various

**[0:05:04] Emma Wu:** linear regression models that are out there? So linear regression is the fundamental machine learning algorithm to predict an outcome variable based on one or more predictor variables, so X and Y. And it assumes a linear relationship between the predictors and the outcome. And the model aims to find the best feed line that minimizes the sum of square differences between the actual and predicted values. But sometimes we do find that using very simple models can cause some problem.

**[0:05:34] Emma Wu:** For example, when you're training your model, there is a situation that the model gives accurate predictions for training data, but not for the new data. It's like at school, you memorize the answers to each question without fully understanding the underlying concept. And this is what we so-called overfitting problem. So sometimes we do want to prevent this overfitting problem by introducing a point called regularization. So a regularization is like you add penalties to the loss functions.

**[0:06:03] Emma Wu:** And the loss function is like a scorecard to minimize the least squares. of your model. There are several variants of linear regression, Lusso, Enrage, and ElasticNet.

**[0:06:15] Eloise Goulder:** So just to stop you there, Emma, you referred to linear regression there as a machine learning algorithm, which I think is fascinating in itself because many of us would have learned linear regression through statistics and would never have thought of it as a form of machine learning. So I think your comment there is a really helpful way of showcasing that machine learning is, in a way, just the processing using machines of statistics.

**[0:06:42] Peng Cheng:** Yes, so actually you bring up a very good point. I would say historically, statistics have always valued parsimony, which means to have as few parameters, as few coefficients as possible in the model, partly because computational power was not there. But in machine learning, we find actually the perception is quite different. If we have a lot of features, some of them turn out not to be statistically significant, but still have some sort of value that conforms to our intuition, there's no harm in leaving them in the model because we think that ultimately they will help without a sample forecasting.

**[0:07:21] Peng Cheng:** So I think there's quite the big difference between statistics and machine learning culturally in how they treat kind of excess parameters. And thanks to the computational power, we can afford to keep these parameters in our model, both in training and in inference.

**[0:07:38] Eloise Goulder:** That is so interesting, this word of parsimony. And I definitely felt that we were trained in statistics about minimizing your number of variables or coefficients. And we were all very nervous about overfitting. But I guess your point is that with the more complex machine learning methods, you can get around some of those issues.

**[0:07:58] Peng Cheng:** Yeah, I think nowadays, a lot of people are too averse to the number of parameters. For example, it's totally okay if we have a thousand observations and 30 explanatory variables. There's nothing wrong with that. You're not going to overfit the model behind this kind of size of observations and explanatory variables. But a lot of people still end up asking me, is it okay to have 30 variables for a thousand observations? My experience is totally fine. So yeah, I think we can learn a little bit from machine learning community.

**[0:08:31] Eloise Goulder:** Fascinating. Thank you, Peng. So Emma, I think you said Lasso, Ridge and ElasticNet are all variants of that basic linear regression model. So can you just define what the differences are between them and what I might know as the basic linear regression model?

**[0:08:47] Emma Wu:** So basically Lasso and Ridge, they both have different penalty terms to be added to their loss function. ElasticNet combines the penalties of both Lussell and Rage so that it can be useful when the data set is large and when there are large numbers of features and the features have potential collinearity among each other. And what do you mean by collinearity there? So collinearity is when your features of your model or X variables of your model are highly correlated with each other.

**[0:09:21] Eloise Goulder:** Yeah, thank you. And the other thing I want to pick up is your use of the word feature, which I know is a commonly used word in machine learning. But as far as I understand, that is the same as the X variable or the independent variable that many of us would have learned in statistics. Is that right?

**[0:09:36] Emma Wu:** Yeah, yeah, it's exactly right. So feature stands for the independent variables. And sometimes you may heard feature selection, meaning that we need to select the variables to be included in the model.

**[0:09:47] Eloise Goulder:** Brilliant. So we've covered linear regression there in some detail. And the keywords there are regularization, penalties, ridge, elastic net, overfitting, and collinearity. So Peng, can we turn to you and can you talk about where you do still find it useful to use these linear regression models in practice?

**[0:10:11] Peng Cheng:** Well, actually linear regression is still the workhorse for quant analysis and it should be the first protocol in most situations. It is easily understood. We can extend it in many different ways. It is easily implementable. So I still find lead integration to be very, very useful. Regularization, I think you have to look at it from two sides. One, as I mentioned, it's actually not very often you'll run into a system where you have so many explanatory variables that you're going to overfit.

**[0:10:44] Peng Cheng:** But economically, actually, we find regularizations to be oftentimes very useful. For example, you may have some economic constraints. Like if I want to track a stock with a portfolio of stocks, sometimes I cannot have 500 stocks tracking this one stock. I may only be restricted to having 30 stocks. So I would need to use Lasso to restrict the number of my variables. So that's one situation. Another situation is actually we find that in our analysis, shrinking coefficients towards zero, just for whatever reason, seems to produce better out-of-sample fitting.

**[0:11:25] Peng Cheng:** So from that perspective, it's almost like you should always use regularization in your linear model. And that's called shrinkage, and it's something we actually use quite a lot in our analysis. So, you know, maybe moving beyond linear model, when should we consider using nonlinear models? I would say that we should do it only if there's a strong economic rationale for nonlinear relationships. And also linear models don't perform well out of sample. Then we should consider moving towards nonlinear models.

**[0:11:57] Eloise Goulder:** So interesting, Peng, that in spite of the existence of so many more complex models that we will go through, So linear models, in your words, are still the workhorse for the quant investing community. And there's still very much a place for linear models. But I think that provides a really good segue into some non-linear models. And I think the first group that you identified, Peng, was random forest. So Emma, could we turn to you again here?

**[0:12:23] Eloise Goulder:** And can you define what you mean by random forest? Sure.

**[0:12:28] Emma Wu:** So before we go to the random forest, let me define the decision tree first, because random forest is essentially a collection of decision tree. So decision tree is like a flowchart that breaks down the problem into smaller choices and each leading to possible outcomes. So random forest is like a collection of decision tree assembled together. An ensemble term that we mentioned here is a general approach to combine the predictions of multiple individual models, allowing multiple models to compensate for each other's weakness.

**[0:13:02] Emma Wu:** And this can lead to an improved generalization and create a more robust and accurate final prediction. There are a few variants of ensemble learning, but I will focus on begging, which is what Random Forest uses and explain in simple words. So imagine you are a PM and you want a group of analysts to help you analyze a large data set. There are a few ways to assign the task. So you can assign each analyst a portion of the data or a subset of the data and ask them to analyze independently.

**[0:13:35] Emma Wu:** And each analyst builds their own predictive model based on the subset of the data. At the end of the day, you combine the predictions to make a final prediction. And this is bagging, which is simply combined predictions from multiple models trained on different subsets of the data. And random forest is a bagging algorithm. The predictions from all the traits are averaged or voted upon to result in a more stable and accurate final prediction.

**[0:14:03] Eloise Goulder:** Thank you, Emma. So the ensemble method is this approach where you combine the predictions of multiple individual models to create that more robust final prediction. And I love your analogy about a PM having multiple different analysts and, you know, asking each analyst to analyze different sub-segments of the data and then the PM putting all of that together. I think that's a really great way of helping us understand Ensemble and the Ensemble method. Are there any other machine learning models that use that method that you think we should note?

**[0:14:34] Emma Wu:** Yes. So we have Extra Boost, which is also Ensemble learning, but use a different methodology, which is called Boosting. So boosting is gradient boosting, and it's different than begging. Let's still use the example that you're a PM and you have analysts and you want to analyze this large data set. So instead of assign each analyst a subset of the data to ask them analyze independently, you can actually assign the sequence of your analyst. And you ask the first analyst to build a model and analyze the entire data set.

**[0:15:07] Emma Wu:** And the second write down where is the arrow of this model. And the second analyst can emphasize the error specifically when building up the model by himself. And the third analyst emphasized the error from the second analyst, so on and so forth, until the last analyst. So in simple words, it's just the next model or the next member, when going into the model, consistently emphasized the error from the previous member or the model.

**[0:15:36] Eloise Goulder:** Thank you. So each analyst is effectively learning from the prior, or in the case of boosting, each model is learning from the prior.

**[0:15:44] Emma Wu:** Yeah, exactly. So other than boosting and begging, there's another field under ensemble learning is called stacking. So we still use the PM and analyst examples. Instead of assigning each analyst or a subset or assign each analyst a sequence, you can basically ask each analyst to analyze the full data by themselves. using different models. So some can use linear, some can use lasso, some can use some nonlinear model. And then you assign a lead analyst that you trust most to take the predictions of each models as the input and make a meta model to make the final predictions.

**[0:16:24] Emma Wu:** So this is the so-called

**[0:16:26] Eloise Goulder:** stacking. That's fascinating. Thank you so much, Emma. And I think your analogy with the PMs and the analysts and the lead analysts is really helpful and quite close to home to many of us. And there are so many keywords you mentioned there. You obviously referred to ensemble learning as the overarching principle behind this sort of model. And then you've talked about bagging versus boosting versus stacking. You've referred to random forest and trees, which obviously fall within the bagging category.

**[0:16:53] Eloise Goulder:** You've referred to XGBoost, which falls within unsurprisingly, the boosting category. You've also referred to the stacking category. So I think there's eight or so keywords there that are really helpful to understand. So Emma, in terms of use cases, what would you really highlight in our domain for these ensemble machine learning methods?

**[0:17:13] Emma Wu:** So last year, we developed a systematic tailhatch model to apply FX options selectively in risk of market and we call it a tailhatch late cycle currency model. We find random forest is particularly useful in selecting features and it also outperforming other models when backtesting. So as of feature selections, initially we start with a list of about 30 signals across different categories such as value, momentum, swap rates, and volve factors. And we first incorporating all of them, but we find this is going to increase the complexity of the model and we observe that overfitting an out-of-sample window.

**[0:17:53] Emma Wu:** So in order to make the model easier to interpret, we applied recursive feature selection under the random forest mechanism and in order to filter the most important and relevant factors. Basically, we start with all the signals included, score them by some criteria, discard the least important ones, and repeat the model, repeat the process until a desired number of signals remained. And by using this approach, we ended up using four features that are selected by this methodology.

**[0:18:25] Emma Wu:** As of the model construction part, the model is developed to predict the pin-l of the FX options. Among all of the models we considered, we find that linear underperformed nonlinear models with a relatively higher tracking error and poor explanatory power. and among all of those non-linear models, we find random forest outperformed others after we successfully tuning parameters such as max features to be used and number of trace to be used.

**[0:18:57] Eloise Goulder:** Thanks. Well, that's really great to hear about a use case on your side where you clearly found that the non-linear models from random forest were more powerful than linear models. So Peng, over to you on random forest models. I mean, what's your take and what do you think the pros and cons of these methods are?

**[0:19:13] Peng Cheng:** Actually, Random Forest is quite popular among practitioners. I find a lot of our clients and conversations seem to use Random Forest in their modeling. It's based on decision trees, which I think are very intuitive and are capable of capturing nonlinear relationships. And also in academic literature as well, Random Forest seems to work well. So I think it's something that a lot of people look to when they first start out exploring nonlinear models. but we also think that it has its limitations.

**[0:19:45] Peng Cheng:** So we often get asked the question, what are the best models to use for a problem? And my answer is always, well, the best model is the one that captures the underlying data generating process the best. So if you think about trees, they are pretty good at capturing nonlinear relationships, but only the discrete ones, right? When we make decision trees, generally speaking, the processes are discrete. But in finance, most of the processes are continuous. So it's actually not very good at capturing continuous relationships.

**[0:20:18] Peng Cheng:** To do that, we can use a number of decision trees to approximate this kind of continuous relationships. And that's why random forests are used. But with this kind of compromise, what you have to do is you have to resort to more computationally intensive variations and lose the interpretability. So those are the limitations of random forests.

**[0:20:42] Eloise Goulder:** So interesting. Thank you. This idea that a basic random forest, the decision tree, of course, is identifying discrete categorizations of the data. But you've used the words before that we live in a nonlinear world. We also live in a continuous world, perhaps. And so I can see those disadvantages of the basic random forest. So can we move on to the third major category of machine learning models now? And I think you said the third one was support vector regression or SVR.

**[0:21:09] Eloise Goulder:** So, Emma, can you start by describing exactly what this is?

**[0:21:13] Emma Wu:** Sure. So support vector regression is like you joining a line in the sand to separate two clusters of points as clearly as possible. Imagine you have two types of stock trades, those that's profitable and some unprofitable. They are categorized by certain features or variables such as P-Ratio and Pulse Returns. And ideally, you can draw a line as a function of these features and separate the profitable trace from the unprofitable ones. And this line is known as the decision boundary.

**[0:21:45] Emma Wu:** What if the decision boundary is not linear? We can choose certain hyperparameters such as the kernel function to create nonlinear decision boundaries. And that being said, the core idea behind support vector regression model is to find the optimal line that separates different classes while maximizing the margin or the space between them. And the support vector term stands for the data points that closest to the decision boundary.

**[0:22:14] Eloise Goulder:** And you use the word hyperparameter there. Can you describe what that means?

**[0:22:18] Emma Wu:** Sure. Hyperparameters in machine learning are like settings on a recipe that you can adjust before cooking. Those are not learned from the data but can be set before the training process. And they govern the behavior of the model and can significantly impact its performance. Hyperparameter tuning is another term that involves finding the best combinations of each hyperparameter values. So for support vector regression here, hyperparameters include the choice of the kernel. For example, you can choose linear kernel, polynomial, or radio basis function, and so on.

**[0:22:54] Eloise Goulder:** And presumably, the fact that you're setting your hyperparameters in advance means that in the case of SVR, it really helps to have some prior knowledge of the data and the domain so that you can effectively set those hyperparameters.

**[0:23:09] Peng Cheng:** Yes, prior knowledge helps, but even without knowledge, we can also train these hyperparameters based on other numerically intensive methods. So it's not a problem.

**[0:23:20] Eloise Goulder:** Very interesting. Thanks. So Peng, what use cases would you really highlight for support vector regression?

**[0:23:27] Peng Cheng:** Yeah, so support vector regression is actually one of our favorite methods. The reason is that it's able to capture nonlinear relationships that are continuous, as Emma mentioned, with the choice of appropriate kernel functions. And these kernel functions are able to generate decision boundaries that are nonlinear and yet very well-behaved. So it also prevents overfitting as well. It's a very well-known model. It's been developed in the 1990s, so packages are easy to find, they're easy to implement.

**[0:23:55] Peng Cheng:** So what kind of nonlinear relationships can we capture? Well, we've done a study where we try to predict future returns of equity futures as a function of historical volatility. So if you think about this relationship, we can generally agree that in periods of low volatility, equity returns are probably good, right? But as volatility picks up, then equity returns declines. But if volatility gets to such a high level, it probably represents a very extreme sell-off situation, and the expected return is probably going to get better again.

**[0:24:26] Peng Cheng:** So as you can see, this is a classical example of a nonlinear relationship. And we find that by using SVM, It actually produces much better out-of-sample performance than a number of other machine learning models. So that's a good use case for support back to machine. Now, what's the main drawback? It's that it's very computationally heavy, and it doesn't really scale well to large problems. So as you increase the number of data points, the size, the training sample, the training time increases exponentially.

**[0:24:58] Peng Cheng:** But sometimes, you know, if it's economically sensible, we will go into these more complex models, especially when simpler models don't perform well. That's really clear.

**[0:25:07] Eloise Goulder:** So another huge number of keywords there that you've gone through. So you've discussed the support vector machine, you've discussed the decision boundary, the kernel trick, which could be linear or polynomial, the hyperparameters, which includes the choice of kernel and the regularization parameter, etc. et cetera, and then support vectors and clustering. Well, what a lot of material. I think this is a great point to wrap up part one, and we'll obviously keep going with defining all of these models in part two.

**[0:25:38] Eloise Goulder:** So Peng, Emma, thank you so much for taking the time to go through all of this with me today.

**[0:25:43] Peng Cheng:** Yeah, thank you, Eloise, for having us.

**[0:25:45] Eloise Goulder:** Yeah, thank you, Eloise. Great to be here. Thank you also to our listeners for tuning into this biweekly podcast from our group. If you'd like to learn more about Peng and Emma's work, then please do head to their JP Morgan Markets page by searching Investable AI. Otherwise, if you'd like to get in touch, then do head to our website at jpmorgan.com forward slash market dash data dash intelligence. And there you can always send us a message.

**[0:26:11] Eloise Goulder:** And with that, we'll close. Thank you.

**[0:26:15] Speaker 4:** Thanks for listening to Market Matters. If you've enjoyed this conversation, we hope you'll review, rate and subscribe to JP Morgan's Making Sense. to stay on top of the latest industry news and trends. Available on Apple Podcasts, Spotify, Google Podcasts, and YouTube. The views expressed in this podcast may not necessarily reflect the views of JPMorgan Chase & Co. and its affiliates, Together JPMorgan, and do not constitute research or recommendation advice or an offer or a solicitation to buy or sell any security or financial instrument.

**[0:26:54] Speaker 4:** They are not issued by J.P. Morgan's Research Department, but are a solicitation under CFTC Rule 1.71. Reference products and services in this podcast may not be suitable for you and may not be available in all jurisdictions. J.P. Morgan may make markets and trade as principal in securities and other asset classes and financial products that may have been discussed. The FICC Market Structure Publications, or to one, newsletters, mentioned in this podcast are available for JPMorgan clients.

**[0:27:24] Speaker 4:** Please contact your JPMorgan sales representative should you wish to receive these. For additional disclaimers and regulatory disclosures, please visit www.jpmorgan.com forward slash disclosures. Copyright 2023, JPMorgan Chase & Company. All rights reserved.
