Welcome to my session on using learning for the benefit of the economy. Before I start um with the main presentation, just I'll introduce myself uh really shortly. Um I'm joining you live CG Plan. Super excited to talk about this.Um This is something, this is a topic I'm super passionate about. Um I'm working uh so well. First of all, the lady in the photo is not me, it's just from the Google Search. Uh but she looks uh well, I don't know if you can see with the camera but she looks like me. Well, same hairstyle, same almost similar color and then a lot of in front of me. Um I'm currently leading the dialytic team uh of uh one of the biggest uh Nordic Baltic retailers called Stockman. And I also have a double life as a phd student. Um And this talk of today is inspired by my experiences as a phd student. I thought about presenting something related to my day job. Um and how data learning can be used in retail. But I do believe that um the economics applications of machine learning are maybe more impactful and they do get unfairly little attention. So I do hope this will be very interesting for you. So, let's see what's on the menu today. I'll start with the motivational. So I, I have this talk today.

Um, and then I'll, um, talk about the coolest project uh relevant for the topic and then just a little bit about what is next in the field. So, a little back story how I came to talk about this today. Um Just if I start, I'll go back because I'm sharing the chrome caps. So I'll just go back to make sure you guys see there is no comments that people don't see. Um Just a minute. Yeah, I think there's no problems. All right then. Um So uh I got involved in tech many, many years ago. I think it was 50 years ago. I started my bachelor's, I did my bachelor's in computer science and then I worked as a web developer. I was one of the two only females in the 20 person company as was usually the case back then and also somewhat now. Um And then I thought, well, this is too much tech for me. I fell victim to this social expectations about what a woman should do. I moved the way back and I see it for my masters and I found the economist to be a little bit boring, but it happens. So that for a project, I would, I had to do some coding for like for an Econometric model and I really enjoyed the coding bit of it and then applying it to economics. And all right, I'm going to continue with economics.

It has to involve some kind of a coding. Hence, this led me to my phd topic which is using machine learning for um economic modeling. So I've been doing this uh phd for six years now, I'm going to finish this year. But over time, I've came across a number of studies where for example, the academics use machine learning to um for example, economic policy making more, some key important economic relationships, trying to predict recessions and so on and so forth. And then the findings from these studies were used by policymakers, politicians, governments, central banks and uh et cetera. But what struck me is that unless you're in academia or part of a research institution or just someone that's very, very curious about these topics, you're highly unlikely to even learn about such studies and their role in the economy. So if this was a face to face station, I would have asked you how many of you have heard, for example, any of the central banks or private banks using ML for for example, macroeconomic modeling policy analysis. For example, you guys know what IMF the International Monetary Fund, uh how IMF use this machine learning?

I think there might be something like that but probably not, not so much or not as much as they should be because the reason that the research academia Central Bank stays mostly within the circles, academia policymakers. But if I ask you how many of you have heard about, you know, ML applications in automobile computers, mobile phones, Instagram filters, I think, you know, maybe all of you would raise your hand. So so much emphasis include in private sector progress using ML uh machine learning.

For example, the voice recognition CD, Alexa face recognition for Instagram, Snapchat filters, self driving cars, robots, etcetera. Uh So we all know about that and this is super uh exciting. I myself use many of these um things uh they bring amazing benefits to us. But I do find that, you know, central bankers, economies, policymakers have no impact making, improving, improving our overall life quality through, for example, controlling for inflation, the fact that we're you know, able to buy, for example, for mobile phones or computers, et that's because for example, someone out there is trying to target the inflation control it so that we don't have the prices do not have to rise as crazy and I'm not talking about the recent month.

Um so and that's the reality. So in the slide, you see a screenshot from the news I took this I think yesterday and you can see this one from BBC, one from economist, one from financial content. There's really nothing that that says you know, central bank did this, I don't know Bank of England did this or that we do see them very rarely but most of all the news and this is from the business and tech sections, most of the news, you know, we can see, for example, uh Elon Musk, asking the workers to start working remotely or for example, what is Google's new scheme on scale, et cetera, et cetera?

So nothing really about what I'm talking today. So then what, what is the end of this talk? Um This, well, this is not specific to women only, but I still hope because this is a woman in tech event. I still hope increase the awareness and inspire women. And maybe by listening to the presentation, you can inspire to involve more into economic research. Plus, I mean, there's also another reason the big guys at PWC uh hope, you know, it's one of the big uh old companies. They say that machine learning in economics can increase productivity, the economic productivity by up to 14% by 2030. So that's an incredibly high number. If we're able to really increase the productivity by that much, that's, that's amazing. All right. So enough of the rambling now, let's go to the part of the, the dirty part. Um Let's see examples of how ML can be used in economics. So essentially the tools are the same in uh the tools you would use uh in economics are the same as in almost in any other uh industry. Um So some of you may be familiar with this, but I, I will try to highlight the applications in economics and bring some really cool examples. And I'll try to keep this talk as much non-technical as possible because I, I didn't know what kind of audience to expect.

So I'll, I'll, I'll keep this non technical. Um And if you have any questions, I'm happy to answer uh at the end of the questions because now I don't see any of the questions if there would be any, I'm only seeing my uh presentation. Uh So the main areas where machine learning can be used in economics are, for example, for forecasting and prediction. So we can use uh various ML two such as rich regressions, random forest, uh K nearest neighbor methods to try to predict inflation rates, for example, unemployment rate, uh try to predict recessions. And I'll have an example in the next few slides uh demand. Uh So um ma many more things. And there's a number of that show that using ML two instead of traditional linear regression, for example, uh one can get the superior productive accuracy, which is what we're aiming for. Right? Then we can use um MLS uh sentiment analysis and natural language processing methods. And this is something that I'm doing for my research. So I'll have it on that later on too. Um And um this, this is an ever increasing research area. There's a number of papers that now tried to use textual data or for example, voice data to try to predict something or analyze uh some information in real time and et cetera uh then image processing and computer vision. So this was the most surprising for me.

Uh 1 may think how can computer vision, for example, have the economy, right? But we all know for example of the satellites that are flying um around the earth, right? And apparently there is a number of uh researchers that use the satellite data, particularly the data on the light. So the electricity light, that's what I mean and try to predict the GDP growth. So that's super surprising, right? At least I wouldn't have that problem. Um And the reason is that in a lot of Sub Saharan African countries, this GDP measure is either not available or it's not accurate. So this way you can basically have lots of data on the light usage and by tracking the growth of the light usage, you're e you're able to estimate the true economic growth. Then there are other studies, for example, using under 60 Panora image of suitcase. So those for example, we see in Google maps to find an empirical connection between physical appearance of a town and behavior and health of its inhabitants. And then this can again be used for policy analysis or changing something in the healthcare system. Then other areas, for example, would be process automation and optimization. Here we can apply any or all of them. Uh ML tools uh to try to say, for example, labor costs, increased productivity growth of the economy.

Uh We can also try to replicate the human behavior um for, for, for controlling, for example, supply shock policy analysis. So for example, by using some ML tools to try to replicate human behavior, we, we would know if we implement some policy or for example, if the next hit or something like that, how would people behave and therefore adjust the reaction of the government accordingly or just prevent some of the reactions.

So this was just the overview of what kind of ML tools you can use and now I go to more specific examples. All right. So the first one is the Billion Prices Project. Uh In my opinion, this is one of the coolest projects that come, came out of academia and this was an inspiration for my own research. So billion Prices Project or BPP, I'll use it interchangeably. Um It was founded by two mit professors Roger Roberto Rabon and Alberto Caballo.

I had the pleasure to meet Robert Roberto Rigon last year. He, he, he's, he's amazing. Um So back in 2000, early two thousands, they were saying how to measure the inflation better. And the reason there is better because for those of you who don't know the most common way nowadays to measure inflation is through survey data, for example, and the survey data are built through people going out in the streets and asking the random uh sample of population. What do you think the prices are now? What do you think the price, this will be the next 12 months? Um So that's the survey method. And also one way is to go manually, prices different sources like supermarkets, I don't know, um car dealers, et cetera, et cetera. So all of these different aspects are then joined together and using some uh formula, the inflation rate is calculated. But the problem is that the survey data is not, you can't really go out in the street and ask customers their expectations or thinking every day, right? So it's not really available with high frequency in the best case you get once a month. Um it it it's just uh you don't have too much data, right? So what they realized if they can get organic data, which is data that is generated without the intention to generate the data, they can get the best measure of inflation. And indeed, they started collecting data from the online supermarkets.

And now, you know, back then there weren't so many. But of course, nowadays, most of the supermarkets have also their own uh versions. So they started with supermarkets and then they moved to other retailers too. And essentially what they do is by Wero they get the prices from different uh items from different online supermarkets. And for all of the countries they work with um and then build the inflation measure from there and then input it to the same model that the central bankers used to uh calculate the inflation. So they just change the traditional measures to the this new measure. And after that, that they do and they use some machine methods to do now testing which is kind of trying to predict inflation in real time. So basically, they're able to predict inflation on a daily level as opposed to uh the traditional methods when you are able to only get it once a month. And currently they collect uh daily prices from almost 70 countries, more than 1000 online retailers and actually the number might even be higher. And why is this exciting?

So this is a snapshot I took from Bloomberg from a few days ago, it says with 58% of inflation, nobody in Argentina knows the price of anything. Um And well, well, this is really bad news, the is incredibly high, but I'd say at least Argentinians now know what the true inflation is because this wasn't the case many years ago. So uh one of the, the I told you the BPP was founded by two researchers and Alberto Cavallo's father is from Argentina. So he started when they started collecting the supermarket prices, he started to uh collect the prices from Argentina as well and then build the inflation.

But what they saw is the as they built. So the annual inflation rate they would build was about 20%. Uh sorry, 11 no, no, no, no, I'm sorry about that 20%. But the one that government was communicating was 11%. I got confused there. So the government said the inflation is 11 but the actual inflation was 20 based on the what they're seeing from their method. Um And there is a slide here which shows just that this is just a snapshot from a few years. But you can see the data index for online and offline online being the solid line, offline being the uh dotted line. Um Same for the annual inflation. And you can see this incredible difference, right? Um And well, there was this difference and you may ask, so how do you know which one was correct? Well, the general population would have that there was more than 11% increase in the prices when they were doing, you know, their daily shopping, for example, and the government was hiding the truth. This new measure that the BPP researchers built became a widely used uh measure in the private sector, not not by the government. Of course.

And long story short, there's of course more to this story, but long story short, what happens next is that uh IMF the interest on monetary fund uh puts Argentinian government under investigation for, for manipulating with public trust. Um And the Argentinian government stopped publishing the inflation rates at all for there was a gap in a few years. They weren't publishing the uh inflation rate. Um But then there was a newly elected government in 2018 that started publishing the true inflation statics, which are now very close to what, what the researchers at BBP calculate. And this is exactly the impact I'm talking about. I'm sure I don't know if any of you heard about this, if you did, let me know, but this is if there wasn't this researcher novel approach to economics to try to calculate inflation differently, it would have been hard to prove that this is happening and then population would still be in the dark or maybe it had taken over to uh figure this difference out.

So this is one of the reasons I think this is a super exciting project but also uh important to share with you. All right, let's move. Um Before the next exam, let's see if they're working on next. So um the this uh BPP team continues to publish a lot of papers. If you're curious, you can just go billion prices project and then check the what they're doing. I've included here, some of the coolest things I've noticed and that also use supermarket or transaction data. So for example, one of the things they're doing now is to um because of the all the lockdowns, the consumer behaviors change, right? That what we both changed and therefore you can't really use the old items to calculate the inflation. So they're trying to analyze this impact of COVID on consumer consumption baskets again using supermarket prices or for example, what they're also trying to analyze again, using this same data is how did the US, the ban on the or the tariffs that us place on Chinese imports and exports?

How that affects the US economy? They also uh again, very impactful work. They're trying to see uh during natural disasters for them in Chile and Japan, how do the prices and supplies change? So then the government or whoever is responsible can act on it. And another thing which I think is super cool. They are now trying or they're actually doing it already uh building the true inflation index for Venezuela. And there is a spin off project where they, it's called inflation Verdadera. And they do this for some Latin American countries.

I think Venezuela is quite important nowadays because I think you must have also read the news, there's a lot happening there. Um So yeah, this is uh again, one big impact the team has. OK. The next example, again, very interesting. Um It's about an impact that was worth the Nobel Prize. Um I, I don't know if you've heard about the these two researchers that are two Nobel Prize uh Este Flow and her husband uh Abhijit Bar. I hope I pronounced it correctly. And so they won Nobel Prize in Economics two years ago for their unique approach in tackling poverty, which uses um A I to deploy the findings of the research in the most efficient way possible and their work directly and indirectly influence national and international policy making in the official ways.

I'll tell you some examples. Uh But I'll first I say what they did. Um So they've introduced a new approach to obtaining a reliable answers about best ways to fight global poverty. So it's before um everyone would be like, yeah, let's fight global poverty without, you know, kind of signable actions. So what they did is they, their approach is to divide the issue into smaller, more manageable questions and then try to tackle those and then using ML to try to find the most effective of these uh problems or where they should actually um spend where the government or spend the funding instead of, you know, just saying, oh, we give you this much money and try to solve the poverty.

Uh They use a um ML to uh ML algorithms to effectively analyze possible solutions and their impacts so that you can focus or the government uh resource allocation can focus on those. And for example, as a result of one of their studies, they did many. So it wasn't just a study. Um But as a result of one of their studies, more than 5 million Indian Children have benefited from effective programs of remedial tutoring in schools or for example, the research helps to get heavy subsidies for preventive healthcare to be introduced in many countries, which there before.

And these are also of course for countries that are more, have higher rates of poverty uh or one of the other examples which again, um and the reason I'm bringing these examples to show how different uh or how many ways their approach can be applied. So another example is for example, they designed this field experiments that show that on the whole farmers in this uh poverty prone countries purchase more fertile if it was offered to them at a small limited time discount earlier in the growing season, rather than giving them more discount, but later in the season.

So it seems very simple. But up up up until then, no one has thought about this. So if you um give more fertilizer to the farmers, um obviously, the result is that you have more um more output. No one thought that, you know, you had to do this early on in the growing season rather than later by giving them very big discount. And uh one more thing is that uh actually um Esther, she was uh the youngest person and only the second woman to receive the Nobel Prize for Economics. But I think it's incredible that she managed to, that she inspired a lot of people including me. Um that not only you can do something valuable but also that recognized for that. All right. Next, moving on. Next. Let's see how machine learning can be applied in central banks. Um Nowadays, there's still there's a wide range of high quality traditional data available. But still these data sets are released with L A. So you don't have a real time data, it can take few days or a few weeks or several months after the reference period. So for this reason, the central banks have been looking at ways to exploit um timelier data and employ more sophisticated methods to enhance forecast accuracy. That um when forecast aury for policy making obviously, and um central bank have stepped up their efforts.

Um sys like even more after the financial crisis of early two thousands. Um So they're all the time trying to find these new ways uh new data to do that. And nowadays available, this uh no data set or alternative data set. It can be both on structure and structured data. Uh I don't know how many of you are familiar with what structure data is, but I'll bring some examples in any case. So the structure data is, for example, the transaction data, payment data from private banks that is able to provide critical real time information, for example, on a consumption economic activity. Uh and that can help to, for example, track uh GDP and that's what a lot of Eurozone countries are now doing. They're using this transaction data to forecast real GDP and structured data is. On the other hand, is the more messy data So such as news articles, images, social media data or just internet altogether. So these, these are a bit more novel sources of data. Uh And these are also being exploded a lot nowadays. Um There will be some example uh soon including my own research. Um And so in addition to this data, then they also are able to use NOL ML tools. Well, they're not really novel, they've been there long, but for economists, most of them are novel.

And nowadays they're able to use, for example, regression trees, a lasso new network or machine to exploit the potential insights from these data sources. So it's, it's actually two new things. They're able to use the traditional uh the novel data and also the ML methods. And uh well, these have been proved to, to improve the forecasting accuracies and help to model key economic agencies where it was a bit harder previously, uh or assess for example, business cycle development and many, many more things. And so um you may ask, can the next global crisis be prevented or predicted? Um Not sure if it can be prevented as such. It, I think it depends on the scale predicted for sure. Um And uh I told you uh after, after the first financial crisis or not the first, but I think first, maybe in our uh in my generation lifetime, the central banks have stepped up their efforts. Um But also COVID has accelerated this trend even more. And nowadays, for example, uh they use a lot of textual data which allows to construct proxies of other variables such as senti mode or are centered uncertainty, there is no uncertainty index in the traditional data. So you build it from text, for example. Uh And then you can use this to check for macroeconomic fluctuations. And this can also serve as early indicators of financial crisis so that the banks would have time to prepare. So maybe not necessarily prevent, prepare.

Um And then some standard uh standard linear models, they work well in con but whenever there's the big shift in the economic outlook, for example, also with COVID, we saw that um then a male models perform much, much better to filter out noises. Whereas traditional ones, um I have some other examples here. Uh Besides um uh what, what I already told just some very random examples. But so you can see there's even more use cases on of this. It's not just limited, but it's not limited to poverty, government or inflation, but there's many more applications and it is only a very, very small subset of the potential applications. Um So one of the um other examples would be now casting uh I talked a lot, a little bit about this but I thought it's worth putting out here. So a lot of uh central banks in Europe um and also ECB the European Central Bank, they're using textual data nowadays to now cast the GDP unemployment, business cycles and so on and so forth. I've also seen, for example, um, Google searches being used to predict unemployment.

I've seen studies in Finland in Germany in the US. Um Spanish uh central bank uses search for private consumption, for example, private banks. Um They also use, well, I think with the private banks, it may be a bit more familiar but still, uh with the private banks, they use machine learning to try to identify um liquidity, predict non-performing roles or defaulted customs. Um or in general, it doesn't have to be for one bank only. They can try to predict uh the default loan default for the Euro level in general. They can also use um the natural language processing um tools to try to summarize loan documents such as like documents, they all they're too long or for example, extract customer sentiment to then try to act on them start ups. Not surprisingly the uh use ML but um the one example I brought is about using it for, for a wider benefit of couple of years ago I met start up. Lastly, I'm not sure if they're still rating, I couldn't remember the name, so I couldn't Google them. But um there was a start up that was using the voice recognition. So when a customer called to the private bank, they're using the uh customer's voice and analyzing it to try to predict if the customer will default on or not. Which again, I find it very interesting.

I, I wonder if they, uh I check it out. I'm sure even if they're not doing it, there will be someone else does it or something similar. Uh Another example would be for the labor market. For example, internet based data can be used to assess the tightness of the labor market or even housing market. Um in the Euro area. For example, there is a measure of labor market tightness uh based on the number of job, based on the job postings. Um And our Irish government, for example, uses this measure. We can also use ML to understand the housing prices. And in general, of course, the state of the economy because they're closely related. Um For example, uh there was a study in Italy that found that metrics based on web script data from an online portal for real estate can be a leading indicator for housing prices. Hey, now it's time for my own contribution. Um I'll try to keep it again and no technical and I don't think I have too much time left anyways. Um So what I do and this is just one figure from the whole basis, but I'm trying to extract inflation uh from my. So I'm using uh natural language processing, I'm using textual data. Uh And I'm trying to convert this textual data into quantifiable measure of inflation.

And as a showcase of my success, this is one of the first figures I made in red, you see the Consumer Confidence Index in the UK. So this is a survey based index where people forecasters are uh the survey collector, they go out in the street and ask the people in the UK, what you think the inflation or what is your feelings about inflation? So it's not the number of inflation but just in general, the fe feeling or the confidence about the economy about the inflation about the future. Um and then the measuring red and then there is the measuring green which is the measure I built from the news. So this is not the inflation measure, this is just the sentiment measure from the news. I then use it to build the inflation. But um you can see the similarity in the shape in the trend. There is quite a strong movement in the series. The correlation is about 0.5 which is quite high for this kind of series. Um And it it even um get as high as 0.7 if I shift my series. So the green series to one month forward, which means that my series are able to predict. So the news based sentiments are able to predict consumer sentiment one month before the official um official measures can.

Um so this was like I said, this was the first figure I made and when I saw this, I thought well, yeah, for sure, this is a reassurance that the news based measures can indeed be good indicators for the economy. So if you just have to chat about my research or you can check in my linking, I have some of my publications there. Um If you're curious, but it's a bit more technical. All right, the very last slide, let's finish up this uh talk. Um So what is next? Um Well, the, the reality is we don't know exactly what is next, but there is a list of predictions by Susan Ay. She's an economics professor in MIT and she had, she has a number of influential papers on applications of uh machine learning economics. And um she has much more predictions, but I included those that I agree with and I changed some of them. But basically what's going to happen now is that more and more research and even more applications of machine learning will be um seen for the purposes of uh the economy. We will also see extensions of these ML methods to try to account for considerations such as fairness, manu manipulable and interpret ability. So imagine if we're able to try to get a measure of government manipulable using machine learning that would be super cool.

Uh And the machine learning method will also be developed even further to adjust or to accommodate to this rapidly changing economy, which is what we see nowadays. Um We also predict that there will be increasing in interdisciplinary research and teaching which is something we also see now.

And my, this the last one is more or less my own prediction is the economist will act more as engineers. Um So not just theoretical economist, but they will be more like engineers and then they will engage with the government, with the firms to design and implement policies in the digital envi environment. So the I think the economy's role will increase. Um And the more people start valuing the job they do. All right. So this is from me. I hope this was interesting. Um I hope you learned something new and I hope this was also inspirational. I'm happy to answer any questions you might have either in the chat or also then in the private. Let me see if there are um some comments. Yes, so much knowledge. Unfortunately, it is locked away in academia. Yeah, I, I totally agree. I agree also with the ethical issues, Kara, that's, that's true. That's why I think that that's why I thought it's important for me to do this talk today rather than talking about something else. All right. Um We have five minutes left. If there's any questions, happy to answer or comments or feedback. Um How would you summarize it in three hashtags uh the, the, the, the talk um the benefits. OK. Uh Oh God, that's a really good question uh I'd say and hashtag underrated hashtag more attention needed.

Um hashtag let's promote research. Oh, I'm about to use multiple. Hashtags All right. If no more questions, I feel free to leave guys if you don't have any questions, uh I'll stay until the end of the session be around. You can text me. I think in the networking session there is this feature or just separately or just connecting the linking. Uh oh And we do have a question. I'm interested in what you mentioned about falling victim to socialization for movie for women and moving out of computer science, you eventually found your way into another male-dominated area. Uh Well, surprisingly economics, not male donated, at least not on the like the student level, like masters, student level or then a phd level, it's not as male dominated, maybe on higher levels. Like, um, I was in a phd conference a couple of weeks ago and it was full of, you know, very old men and, uh I, I noticed that there were tables of 10 people, uh each and about 15 days and in each table there were only like one or two women. So, yeah, you're right. I think it's still dominated. But, uh but in student level it's not that donated. But I think, you know, as I grew, I just realized I shouldn't care if it's male or women dominated. I just have to do what I like.

So, but then I was, you know, very young in my late teens. Um, and I just grew, I guess, and so and you're asking how you find this part? Well, like it's just, I don't pay attention anymore. I do what I want to do. Thank you and thank you both for the comments and questions. Well, uh thank Marta for being here. I, I don't think we'll have more questions. Oh, no, we do. Thank you guys. You made me think about all the ethical issues surrounding use of ML in economics. Yeah. True. And if you're, if you're curious um not exactly about ethical issues but the the Susan Athens, she has this uh really nice few papers she wrote, I can share with your curious just how mail can be used. I think the ethical issues is something we imply ourselves by seeing all of this. Um I don't think anyone talks about this really. Uh Yeah, I also hope so and I think we're somewhat seeing that maybe nowadays because you know, with everything happening with COVID, with um the war with just the general, the markets, we are all see the situation in the stock market with the prices rising, et cetera. Um I'm somehow feel of course, it's just my feeling we never know what the big guys out there are thinking. But I think uh we are seeing that the response at least is more real time and more timely.

And that's thanks to all the data we have available, which we didn't have back in 2018, 2008. Thank you too. For participating and I'm wishing you a lovely rest of the conference and uh happy to connect with you over linkedin or over the networking sessions. Bye bye.