ARNIE in Action – OeNB’s Stress Test Framework Including Climate Risk - MATLAB
Video Player is loading.
Current Time 0:00
Duration 24:27
Loaded: 0.68%
Stream Type LIVE
Remaining Time 24:27
 
1x
  • Chapters
  • descriptions off, selected
  • en (Main), selected
    Video length is 24:27

    ARNIE in Action – OeNB’s Stress Test Framework Including Climate Risk

    Martin Guth, Austrian Central Bank (Österreichische Nationalbank)

    The Austrian Central Bank (Österreichische Nationalbank, or OeNB) conducts stress tests to assess the resilience of the banking system and to analyze potential drivers of systemic risk, complemented by more specialized exercises (e.g. climate risk). Martin Guth, a stress test analyst at the OeNB, talks about ARNIE, the bank’s stress-testing platform, which was created using MATLAB®.

    Published: 14 Nov 2022

    Yeah. Thanks for a nice introduction. My name is Martin Hafner-Guth. I work for the Austrian National Bank as a stress test analyst. And today, I want to talk about our stress testing software, which is called ARNIE, which is a fully MATLAB-based software, which we used for bank system, solvency stress testing, and also since 2021 for Climate Risk Stress Testing. Oh, yeah, and before I start, the usual disclaimer applies that the views expressed from me, those are mine alone and do not reflect the views of the Eurosystem or the Austrian National Bank.

    OK. But now I can start. And before getting into the actual presentation, I just wanted to give you some key facts about the Austrian banking system. So the total assets amount to roughly 1,200 billions in euros, which is quite large compared to the overall size of the Austrian economy. And of these total assets, we have a quite substantial exposure in the central, east, southeast, eastern European region-- so basically, everything east of Austria-- which amounts to roughly 280 billion euros.

    And you can see that in the map here on the right side that, besides Germany, are the largest exposure is in the Czech Republic, followed by Slovakia, Romania, and so on.

    We also have also quite high number of banks. So there are a total of 376 banks on the highest concentration level. And this comes from the so-called life asset sector, which is a banking sector, which uses an inverse consolidation scheme.

    And of these 376 banks, there are six significant institutions which are directly supervised by the European Central Bank. There is one important euro area, SI-Sub, which is a daughter from the Italian UniCredit group, and then the remaining 369 less significant institutions.

    And in the last years, we have seen an ongoing consolidation process of those LSIs to reduce the number of banks, given larger pressure on the profitability side. And since 2015, there have been over 150 mergers in total.

    But now, coming to ARNIE-- and you may be wondering now why lovely Arnold Schwarzenegger is waving to you. So the myth has it that 10 years ago when my colleague started to work on ARNIE, they first came up with the acronym ARNIE because they wanted to have a fancy name for their model-- something that is easily recognizable across Europe in other stressing departments, which can be also related easily to Austria. So they came up with ARNIE, which now stands for Applied Risk Network and Impact Assessment Engine.

    So the agenda for the next couple of minutes will be that first off, I want to give you a short introduction to our stress testing history. So the models that we have been using and the extensions we have implemented so far. Then I want to give you an overview of our generic stress testing model and how we think about bank-level stress tests.

    Then, I want to give you some technicalities about how ARNIE's actually implemented and how it works. I'm also going to talk about results of the solvent stress test and how we use these results in the supervisory follow-up with colleagues from other supervision divisions. And then as a last point in my presentation, I'm going to talk about the Climate Risk Stress Test.

    OK. So coming to our stress testing history, the Austrian National Bank actually started in 2003, implementing the first stress tests, which was somewhere around the time also the IMF started stress testing and published the first papers on stress testing. And these first stress tests have been Excel spreadsheet-based calculations.

    And then in 2006, my colleagues back then wanted to have something more automated, something that could handle more data more easily and would be more easily executable. So they implemented the so-called Systemic Risk Monitor, which was also back then MATLAB and MATLAB-based software, which just operated on one period.

    And the focus of the SRM was credit risk, market risk, and interbank network contagion effects. In the subsequent years-- so specifically from 2008 to 2012-- the SRM was extended to cover a multi-period-- so multiple periods-- and to become more of a macro testing tool. So the focus was to assess how certain macroeconomic scenarios would affect the whole of the Austrian banking system.

    With 2013 and an IMF submission in Austria, the SRM was discharged, and the work on ARNIE has begun. Because the SRM somehow came to its limits in terms of extensibility, so my colleagues back then wanted to set things up new.

    And ARNIE, yet again, is a model-based software. And it fully follows the methodology outlined by the European banking authority. We cover the entire profit and loss positions of each Austrian bank across different risk modules. And I will come to that later. So credit risk, market risk and so on.

    But we also have add-ons for network contagion just like in the SRM, but also now solvency liquidity interactions and specific risk types to the Austrian economy and the Austrian banking systems. Since then, we always try to add something new to ARNIE and to make it better. So for example, in 2020, during the corona crisis, we developed a corporate insolvency model to assess the impact of the fiscal mitigating measures during the COVID pandemic, which was, again, a model-based software. And it fully integrates with ARNIE.

    In 2021, as mentioned before, we called it the Climate Risk Stress Test, which also fully integrates into ARNIE. And it will come to that, that we use all the different kinds of models to assess the impact of the carbon pricing schemes. And since 2021, we currently work on an extension to work on a dynamic balance sheet, which is, in our view, the next big step in banking system stress testing.

    OK. So how does our stress testing model look like? In essence, there are these four big blocks. We start with exogenous shocks. These are defined either by the ECB or by our own in-house economic models. So those exogenous shocks are macroeconomic models.

    This is a narrative structuring-- what we think stress for the banks could look like. So this year, obviously, a big factor was the war in Ukraine, rising inflation, rising interest rates levels, and so on.

    These shocks are then handed over to satellite models, which translates the country-level shocks to bank-level risk parameters. For example, probability of defaults and loss given default rates. And you can see by the coloring here, which boxes are outside and inside of ARNIE. So the exogenous shocks obviously are outside.

    The risk factor distribution is then somewhat already inside ARNIE. And then fully inside ARNIE is the loss functions. And the loss functions are basically simple profit and loss models and expected credit loss models, where we calculate for each risk factor.

    And for each bank, we calculate the full profit and loss statement and come up in the end results in terms of common equity tier 1 ratios. And the nice thing about ARNIE is that, given our very granular database and the loss functions being calculated on a bank-to-bank basis, we can track each profit and loss statement throughout the whole stress test horizon.

    And then on top of the results, we do have different feedback effects, like I've mentioned before. So on micro feedbacks, you can think about contagion effects. For example, when a bank falls below certain thresholds, this would lead to a default, which would then have knock-on effects on other banks. But also, the interactions between liquidity and solvency, cost of funding, have potential feedback effects.

    And on the macro feedback, this is something we're currently working on. So feedback effects ranging from a reduction in bank lending could result in an even more severe macroeconomic scenario. And this scenario would then lead to higher probability of defaults.

    So coming to the actual calculation in ARNIE, I first want to give you a view of how we handle data. And again, the coloring still stays the same. So we start with raw data, which comes in databases, but also just flat files which contains bank metadata, supervisory reporting data on the banks, but also our central trade register, which is taken by an R routine. And we use R in that sense because we have a quite powerful R server in the International banks, which is used to manage these huge data quantities and loaded into an Oracle database, which we call input data.

    And from then on, everything is handled in MATLAB, meaning that we take the input data, give them some transformations, and provide it a generic structure, which is, again, in the Oracle database but in a different table, which we call ARNIE data pool. And what we achieve with this data pool is that we have one table for the different variable categories.

    And in this data pool, we store historic, adjusted, and projected values. And with every year and with every run of ARNIE, this data pool is extended. Hence, we have, by now, a very large history of different data pools, which is quite handy when it comes to backtesting our software and to track our different results and different permutations within ARNIE.

    Now, from the ARNIE data pool, we load the data into the actual ARNIE calculations routines, which then, as described before, will calculate all the profit and loss statements. And the results are then stored back into the data pool.

    Also, a nice feature about the data pool is that with everything stored in there, we can easily reiterate our calculations and still start from the same starting points. But we also can easily go back in time and redo a former stress test, which the data pool guarantees us that we come up with the same results.

    Also, what I should mention is that we also use Bitbucket and cheat tab to recognize our code. Hence, going back in time not only means taking the data from, let's say, one or two years ago, but we can also easily use the actual code we have been using to come up with the same results.

    Now, how does the actual calculation work? And we I don't want to go into details of every box you see here. But just to give you an overview of the different modules we use. So I've also already covered what we call micro to micro. So the translation from the macroeconomic scenario to bank-level risk parameters, specifically PDs and LGDs.

    In credit risk, we have an expected loss few calculating credit losses, but also credit risk exposure amounts. We also track foreign currency loans, which are quite important for the Austrian bank system.

    On profit and loss side, we have modules calculating net interest income, net fee, and commission income, the trading income. And we also have a quite important module calculating the risk from participations, which is quite important for the life asset sector, which I mentioned before, with the inverse consolidation scheme because they have high equity stakes in banks within their own sector, but also high equity stakes in companies within and outside of Austria. And the feedback effects, I already mentioned before.

    Now, coming from all these theoretical parts and technicalities, I wanted to give you also insights into what the actual results would look like. And as mentioned before, so this year's stress test is very interesting, given the current macroeconomic outlook. However, sadly, I cannot present the results of the shear stress test because they are not publicly available yet. But it will happen in the next couple of weeks.

    What you see here is other results from last year. And in last year, there were scenario assumed an ongoing pandemic and supply bottlenecks and other effects caused from an ongoing epidemic. And what you see here in the chart is the set 1 ratio plotted for the Austrian banking system.

    The solid line is the adverse scenario. And the dotted line is the baseline scenario. And what you can see is that the aggregate impact in the adverse scenario is 5.1 percentage points. So from 16.1% to 11% of set 1 capital. And this is roughly 22 billion euros, which would have been lost in the adverse scenario.

    However, given that the Austrian banks, or the Austrian banking system, started at an already quite high 16.1%, we do think that the Austrian banking system shows a solid risk-bearing capacity, even in such a scenario. So I won't go into much more details which would just take too much time. But what I can tell you is that over the years, we've seen that credit risk always remains the biggest risk driver, followed by reduced net interest income, and also the participation risks within the life asset sector.

    Now, what do we actually use these results for? Because it's always nice to calculate these, but there need to be some usage, some supervisory follow-up. So there are multiple paths the results are used. The most important one is the so-called Pillar 2 guidance. So we hand over our results to the colleagues from the offsite supervision and to the Financial Market Authority to calibrate the Pillar 2 guidance on a bank-by-bank basis.

    And for this, the colleagues get for each bank a roughly 25-page-long document, which is automatically created for each bank containing the bank-specific results and also bank-specific text blocks explaining to colleagues how to interpret the stress test results.

    Also, we use it for the so-called Institutional Protection Scheme Fund, which is, again, in the life asset sector. And the Financial Market Authority needs to approve this IPs fund on a yearly basis. And our stress test results are used to assess the floor of the fund, given the stress capital ratios.

    On top of the stress test, we also calculate contagion analysis for financial stability and resolution authorities. And the results are also used for further supervisory strategy requests like currently the rising interest rate environment or real estate exposure risks.

    Now, the last part and I think maybe the most interesting part for you, the Climate Risk Stress Test. So the Climate Risk Stress Test was a pilot exercise conducted in 2021. And our main goal was to use our existing OeNB top-down stress testing infrastructure.

    And so looking at this flowchart down here, we already had ARNIE as our top-down stress testing software. And we already had the insolvency model, which is the insolvency model I mentioned before, which was developed in 2020 during the COVID pandemic. And these are the three blocks you see have been created now in this pilot exercise to perfectly fit in these other two models in order to gorge our already-- the existing experience with these models, and especially use our need to really dig deep into the different P&L items of the banks, given this new Climate Risk Stress Testing scenario.

    For these models, we use relatively well-established data sources. For example, we use for the Sectoral Carbon Pricing Model, which I will explain the second; the figural database by the European Commission. The insolvency model uses an Austrian firm-level specific database and well, as explain before, all the users, all the supervisory data.

    In this Climate Risk Stress Test, we focused on transmission risks. We frontloaded carbon price shocks as the main risk driver. And I will talk a bit about those frontloaded shocks although in a second.

    We used time horizon of five years. So normally, our bank solvency stress test operate on the horizon of three years. So we extended it by two years. And we focus on modeling credit risk impact for Austrian foreign exposures with an additional market risk module.

    So I can't go into all the details of all the models, except ARNIE, which I already explained, because that would take too much time. But especially, for the Sectoral Carbon Price Model and the Insolvency Model, I do have two slides in the annex. So I know that the slides will be sent out to you. And I also linked you here to the financial stability report article covering all the modules so you can read up on all the details. And if there should be more questions, you can always write an email.

    But to go through these blocks quickly-- so the Sectoral Carbon Price Model is in multi-regional input output model covering 21 sectors. And we use carbon price shocks as additional tax shocks on those 21 sectors, coming up with cost and turnover changes, given the increase in carbon price taxes.

    These costs and turnover changes are put into the insolvency model, which uses marginal distributions within maze one digit sectors. And that simulates how these costs and turnover changes would affect the solvency and liquidity positions of those simulated companies in these sectors. And given some sort of liquidity thresholds, we then can see how many of the simulated banks would default and become insolvent.

    So from these insolvent banks, we derive sectoral insolvency rates, which are, sadly, just for Austria because the firm level data we use for the insolvency model is just for Austria. However, the Austrian insolvency rates, together with the cost and turnover changes, which we have for all European countries, are then extrapolated to all European countries within these leaking equations.

    And additionally-- so all these parts are basically credit risk-- we have a market risk module, which calculates losses for bond holdings and equity stakes. And basically, the increased probability of defaults and those valuation losses are then used as shocks within ARNIE to calculate bank-specific capital impacts.

    Now, how do our carbon price scenarios look like? So as a baseline, we use the ECB/EBA macroeconomic baseline from the European Environment Stress Test 2021 as a starting point. And we use two scenarios. We use an orderly and a disorderly transition shock scenario based on the NGFS scenarios. So the NGFS is the Network for Greening the Financial System.

    And you can see in this left chart down here the greenish lines are the NGFS scenarios for the carbon price development. The top line is the Net Zero 2050 scenario, which we use as orderly scenario. So you can see this blue dotted line here is the increase in the cap price for the orderly scenario, which is also reflected here in these blue charts.

    And the disorderly transition scenario, according to the NGFS, which is the lower line here, because it's a delayed transition, would just start in 2030 to 2035. However, given that we operate on a five-year stress test horizon and not a 35-year horizon, we just decided to take this incline and move it to the front, which is here displayed as this red dotted line.

    And you can see also here in the red bars in the right chart basically, having now an orderly transition scenario where the carbon price path would increase from roughly 40 euros per ton of CO2 equivalent to around 140. And in disorderly, it already starts at nearly 140 and goes up to over 250 euros per ton of CO2 equivalent. And these carbon prices are modeled as an additional tax on the already existing carbon price taxes in all European countries.

    And the last point on this slide is that, as I mentioned before, the sectoral carbon price model is modeled for a setup for all European countries. But you also have implemented a Carbon Board Adjustment Mechanism, meaning that all the goods that are imported out of the European Union are also-- they get also this carbon price tax, which follows the European Commission.

    Now, my last slide. Again, some results. So what you see here in this chart compared to the solvency chart is the deviation from the baseline. So in the orderly transition scenario, which is the green line, the CT1 ratio of the Austrian banking system would fall by 0.7 percentage points, which relates to roughly 3 billion euros. Further, in the disorderly transition scenario, the CT1 ratio would fall by 2.7 percentage points, which relates to roughly 12 billion euros.

    Now, if you remember in the solvency stress test from last year, I showed you the impact was five percentage points. So nearly double the effect of the Climate Risk Stress Test. And this is due to limited exposure of the Austrian banks to climate-relevant sectors, compared to the way larger CESEE exposure, which, in our macroeconomic scenarios, is hit more often by higher shocks.

    So given these results, we can say that this carbon pricing scenario reviews from the NGFS does not prompt any financial stability concerns. And with that, thank you very much for your attention. And I am happy to take questions.

    Related Products

    View more related videos