Schweitzer Fachinformationen
Wenn es um professionelles Wissen geht, ist Schweitzer Fachinformationen wegweisend. Kunden aus Recht und Beratung sowie Unternehmen, öffentliche Verwaltungen und Bibliotheken erhalten komplette Lösungen zum Beschaffen, Verwalten und Nutzen von digitalen und gedruckten Medien.
State-of-the-art algorithmic deep learning and tensoring techniques for financial institutions
The computational demand of risk calculations in financial institutions has ballooned and shows no sign of stopping. It is no longer viable to simply add more computing power to deal with this increased demand. The solution? Algorithmic solutions based on deep learning and Chebyshev tensors represent a practical way to reduce costs while simultaneously increasing risk calculation capabilities. Machine Learning for Risk Calculations: A Practitioner's View provides an in-depth review of a number of algorithmic solutions and demonstrates how they can be used to overcome the massive computational burden of risk calculations in financial institutions.
This book will get you started by reviewing fundamental techniques, including deep learning and Chebyshev tensors. You'll then discover algorithmic tools that, in combination with the fundamentals, deliver actual solutions to the real problems financial institutions encounter on a regular basis. Numerical tests and examples demonstrate how these solutions can be applied to practical problems, including XVA and Counterparty Credit Risk, IMM capital, PFE, VaR, FRTB, Dynamic Initial Margin, pricing function calibration, volatility surface parametrisation, portfolio optimisation and others. Finally, you'll uncover the benefits these techniques provide, the practicalities of implementing them, and the software which can be used.
Quants, IT professionals, and financial risk managers will benefit from this practitioner-oriented approach to state-of-the-art risk calculation.
IGNACIO RUIZ, PhD, is the head of Counterparty Credit Risk Measurement and Analytics at Scotiabank. Prior to that he has been head quant for Counterparty Credit Risk Exposure Analytics at Credit Suisse, head of Equity Risk Analytics at BNP Paribas and he founded MoCaX Intelligence, from where he offered his services as an independent consultant. He holds a PhD in Physics from the University of Cambridge.
MARIANO ZERON, PhD, is Head of Research and Development at MoCaX Intelligence. Prior to that he was a quant researcher at Areski Capital. He has extensive experience with Chebyshev Tensors and Deep Neural Nets applied to risk calculations. He holds a PhD in Mathematics from the University of Cambridge.
Acknowledgements xvii
Foreword xxi
Motivation and aim of this book xxiii
Part One Fundamental Approximation Methods
Chapter 1 Machine Learning 3
1.1 Introduction to Machine Learning 3
1.1.1 A brief history of Machine Learning Methods 4
1.1.2 Main sub-categories in Machine Learning 5
1.1.3 Applications of interest 7
1.2 The Linear Model 7
1.2.1 General concepts 8
1.2.2 The standard linear model 12
1.3 Training and predicting 15
1.3.1 The frequentist approach 18
1.3.2 The Bayesian approach 21
1.3.3 Testing-in search of consistent accurate predictions 25
1.3.4 Underfitting and overfitting 25
1.3.5 K-fold cross-validation 27
1.4 Model complexity 28
1.4.1 Regularisation 29
1.4.2 Cross-validation for regularisation 31
1.4.3 Hyper-parameter optimisation 33
Chapter 2 Deep Neural Nets 39
2.1 A brief history of Deep Neural Nets 39
2.2 The basic Deep Neural Net model 41
2.2.1 Single neuron 41
2.2.2 Artificial Neural Net 43
2.2.3 Deep Neural Net 46
2.3 Universal Approximation Theorems 48
2.4 Training of Deep Neural Nets 49
2.4.1 Backpropagation 50
2.4.2 Backpropagation example 51
2.4.3 Optimisation of cost function 55
2.4.4 Stochastic gradient descent 57
2.4.5 Extensions of stochastic gradient descent 58
2.5 More sophisticated DNNs 59
2.5.1 Convolution Neural Nets 59
2.5.2 Other famous architectures 63
2.6 Summary of chapter 64
Chapter 3 Chebyshev Tensors 65
3.1 Approximating functions with polynomials 65
3.2 Chebyshev Series 66
3.2.1 Lipschitz continuity and Chebyshev projections 67
3.2.2 Smooth functions and Chebyshev projections 70
3.2.3 Analytic functions and Chebyshev projections 70
3.3 Chebyshev Tensors and interpolants 72
3.3.1 Tensors and polynomial interpolants 72
3.3.2 Misconception over polynomial interpolation 73
3.3.3 Chebyshev points 74
3.3.4 Chebyshev interpolants 76
3.3.5 Aliasing phenomenon 77
3.3.6 Convergence rates of Chebyshev interpolants 77
3.3.7 High-dimensional Chebyshev interpolants 79
3.4 Ex ante error estimation 82
3.5 What makes Chebyshev points unique 85
3.6 Evaluation of Chebyshev interpolants 89
3.6.1 Clenshaw algorithm 90
3.6.2 Barycentric interpolation formula 91
3.6.3 Evaluating high-dimensional tensors 93
3.6.4 Example of numerical stability 94
3.7 Derivative approximation 95
3.7.1 Convergence of Chebyshev derivatives 95
3.7.2 Computation of Chebyshev derivatives 96
3.7.3 Derivatives in high dimensions 97
3.8 Chebyshev Splines 99
3.8.1 Gibbs phenomenon 99
3.8.2 Splines 100
3.8.3 Splines of Chebyshev 101
3.8.4 Chebyshev Splines in high dimensions 101
3.9 Algebraic operations with Chebyshev Tensors 101
3.10 Chebyshev Tensors and Machine Learning 103
3.11 Summary of chapter 104
Part Two The toolkit - plugging in approximation methods
Chapter 4 Introduction: why is a toolkit needed 107
4.1 The pricing problem 107
4.2 Risk calculation with proxy pricing 109
4.3 The curse of dimensionality 110
4.4 The techniques in the toolkit 112
Chapter 5 Composition techniques 113
5.1 Leveraging from existing parametrisations 114
5.1.1 Risk factor generating models 114
5.1.2 Pricing functions and model risk factors 115
5.1.3 The tool obtained 116
5.2 Creating a parametrisation 117
5.2.1 Principal Component Analysis 117
5.2.2 Autoencoders 119
5.3 Summary of chapter 120
Chapter 6 Tensors in TT format and Tensor Extension Algorithms 123
6.1 Tensors in TT format 123
6.1.1 Motivating example 124
6.1.2 General case 124
6.1.3 Basic operations 126
6.1.4 Evaluation of Chebyshev Tensors in TT format 127
6.2 Tensor Extension Algorithms 129
6.3 Step 1-Optimising over tensors of fixed rank 129
6.3.1 The Fundamental Completion Algorithm 131
6.4 Step 2-Optimising over tensors of varying rank 133
6.4.1 The Rank Adaptive Algorithm 134
6.5 Step 3-Adapting the sampling set 135
6.5.1 The Sample Adaptive Algorithm 136
6.6 Summary of chapter 137
Chapter 7 Sliding Technique 139
7.1 Slide 139
7.2 Slider 140
7.3 Evaluating a slider 141
7.3.1 Relation to Taylor approximation 142
7.4 Summary of chapter 142
Chapter 8 The Jacobian projection technique 143
8.1 Setting the background 144
8.2 What we can recover 145
8.2.1 Intuition behind g and its derivative dg 146
8.2.2 Using the derivative of f 147
8.2.3 When k < n becomes a problem 149
8.3 Partial derivatives via projections onto the Jacobian 149
Part Three Hybrid solutions - approximation methods and the toolkit
Chapter 9 Introduction 155
9.1 The dimensionality problem revisited 155
9.2 Exploiting the Composition Technique 156
Chapter 10 The Toolkit and Deep Neural Nets 159
10.1 Building on P using the image of g 159
10.2 Building on f 160
Chapter 11 The Toolkit and Chebyshev Tensors 161
11.1 Full Chebyshev Tensor 161
11.2 TT-format Chebyshev Tensor 162
11.3 Chebyshev Slider 162
11.4 A final note 163
Chapter 12 Hybrid Deep Neural Nets and Chebyshev Tensors Frameworks 165
12.1 The fundamental idea 165
12.1.1 Factorable Functions 167
12.2 DNN+CT with Static Training Set 168
12.3 DNN+CT with Dynamic Training Set 171
12.4 Numerical Tests 172
12.4.1 Cost Function Minimisation 172
12.4.2 Maximum Error 174
12.5 Enhanced DNN+CT architectures and further research 174
Part Four Applications
Chapter 13 The aim 179
13.1 Suitability of the approximation methods 179
13.2 Understanding the variables at play 181
Chapter 14 When to use Chebyshev Tensors and when to use Deep Neural Nets 185
14.1 Speed and convergence 185
14.1.1 Speed of evaluation 186
14.1.2 Convergence 186
14.1.3 Convergence Rate in Real-Life Contexts 187
14.2 The question of dimension 190
14.2.1 Taking into account the application 192
14.3 Partial derivatives and ex ante error estimation 195
14.4 Summary of chapter 197
Chapter 15 Counterparty credit risk 199
15.1 Monte Carlo simulations for CCR 200
15.1.1 Scenario diffusion 200
15.1.2 Pricing step-computational bottleneck 200
15.2 Solution 201
15.2.1 Popular solutions 201
15.2.2 The hybrid solution 202
15.2.3 Variables at play 203
15.2.4 Optimal setup 207
15.2.5 Possible proxies 207
15.2.6 Portfolio calculations 209
15.2.7 If the model space is not available 209
15.3 Tests 211
15.3.1 Trade types, risk factors and proxies 212
15.3.2 Proxy at each time point 213
15.3.3 Proxy for all time points 223
15.3.4 Adding non-risk-driving variables 228
15.3.5 High-dimensional problems 235
15.4 Results Analysis and Conclusions 236
15.5 Summary of chapter 239
Chapter 16 Market Risk 241
16.1 VaR-like calculations 242
16.1.1 Common techniques in the computation of VaR 243
16.2 Enhanced Revaluation Grids 245
16.3 Fundamental Review of the Trading Book 246
16.3.1 Challenges 247
16.3.2 Solution 248
16.3.3 The intuition behind Chebyshev Sliders 252
16.4 Proof of concept 255
16.4.1 Proof of concept specifics 255
16.4.2 Test specifics 257
16.4.3 Results for swap 260
16.4.4 Results for swaptions 10-day liquidity horizon 262
16.4.5 Results for swaptions 60-day liquidity horizon 265
16.4.6 Daily computation and reusability 268
16.4.7 Beyond regulatory minimum calculations 271
16.5 Stability of technique 272
16.6 Results beyond vanilla portfolios-further research 272
16.7 Summary of chapter 273
Chapter 17 Dynamic sensitivities 275
17.1 Simulating sensitivities 276
17.1.1 Scenario diffusion 276
17.1.2 Computing sensitivities 276
17.1.3 Computational cost 276
17.1.4 Methods available 277
17.2 The Solution 278
17.2.1 Hybrid method 279
17.3 An important use of dynamic sensitivities 282
17.4 Numerical tests 283
17.4.1 FX Swap 283
17.4.2 European Spread Option 284
17.5 Discussion of results 291
17.6 Alternative methods 293
17.7 Summary of chapter 294
Chapter 18 Pricing model calibration 295
18.1 Introduction 295
18.1.1 Examples of pricing models 297
18.2 Solution 298
18.2.1 Variables at play 299
18.2.2 Possible proxies 299
18.2.3 Domain of approximation 300
18.3 Test description 301
18.3.1 Test setup 301
18.4 Results with Chebyshev Tensors 304
18.4.1 Rough Bergomi model with constant forward variance 304
18.4.2 Rough Bergomi model with piece-wise constant forward variance 307
18.5 Results with Deep Neural Nets 309
18.6 Comparison of results via CT and DNN 310
18.7 Summary of chapter 311
Chapter 19 Approximation of the implied volatility function 313
19.1 The computation of implied volatility 314
19.1.1 Available methods 315
19.2 Solution 316
19.2.1 Reducing the dimension of the problem 317
19.2.2 Two-dimensional CTs 318
19.2.3 Domain of approximation 321
19.2.4 Splitting the domain 323
19.2.5 Scaling the time-scaled implied volatility 325
19.2.6 Implementation 328
19.3 Results 330
19.3.1 Parameters used for CTs 330
19.3.2 Comparisons to other methods 331
19.4 Summary of chapter 334
Chapter 20 Optimisation Problems 335
20.1 Balance sheet optimisation 335
20.2 Minimisation of margin funding cost 339
20.3 Generalisation-currently "impossible" calculations 345
20.4 Summary of chapter 346
Chapter 21 Pricing Cloning 347
21.1 Pricing function cloning 347
21.1.1 Other benefits 352
21.1.2 Software vendors 352
21.2 Summary of chapter 353
Chapter 22 XVA sensitivities 355
22.1 Finite differences and proxy pricers 355
22.1.1 Multiple proxies 356
22.1.2 Single proxy 357
22.2 Proxy pricers and AAD 358
Chapter 23 Sensitivities of exotic derivatives 359
23.1 Benchmark sensitivities computation 360
23.2 Sensitivities via Chebyshev Tensors 361
Chapter 24 Software libraries relevant to the book 365
24.1 Relevant software libraries 365
24.2 The MoCaX Suite 366
24.2.1 MoCaX Library 366
24.2.2 MoCaXExtend Library 377
Appendices
Appendix A Families of Orthogonal Polynomials 385
Appendix B Exponential Convergence of Chebyshev Tensors 387
Appendix C Chebyshev Splines on Functions with No Singularity Points 391
Appendix D Computational savings details for CCR 395
D.1 Barrier option 395
D.2 Cross-currency swap 395
D.3 Bermudan Swaption 397
D.3.1 Using full Chebyshev Tensors 397
D.3.2 Using Chebyshev Tensors in TT format 397
D.3.3 Using Deep Neural Nets 399
D.4 American option 399
D.4.1 Using Chebyshev Tensors in TT format 400
D.4.2 Using Deep Neural Nets 401
Appendix E Computational savings details for dynamic sensitivities 403
E.1 FX Swap 403
E.2 European Spread Option 404
Appendix F Dynamic sensitivities on the market space 407
F.1 The parametrisation 408
F.2 Numerical tests 410
> 1 412
Appendix G Dynamic sensitivities and IM via Jacobian Projection technique 415
Appendix H MVA optimisation - further computational enhancement 419
Bibliography 421
Index 425
The world of risk analytics has had an ever growing demand for computing capacity since the early 2000s. When one of the book's authors started working in this field, he was asked to work on the CCR engine at Credit Suisse for the new Basel II regulation and IMM capital calculation. At the time, it was the latest big thing in the industry. The IMM-related calculations were one of the most (if not the most) complicated calculations the bank had done up to that point. A few hundred CPUs were bought and installed in a state-of-the-art grid computing farm. The belief was that such grid would be able to do any CCR calculation. However, it did not take much time for the team to realise that more computing power was needed to match the computational requirements of new calculations being requested. Over the years, we have experienced a world in which, regardless of how much computing power is available with the latest technologies, soon it is insufficient to meet the new demands and needs to be upgraded only a few years later.
Indeed, the world of banking, and in particular the business of derivatives, has become a technology race (like many other industries, it must be said). As P. Karasinski says in his Foreword to this book, it used to be about creating and selling the new exotic product. Now it is about computing prices and increasingly sophisticated risk metrics in a prompt and efficient manner - partly as a result of regulations that have become more stringent since the 2008 crisis, partly as a result of the higher standards for risk management the industry has developed. That is where the differentiation between different broker-dealers resides and where the source of profitability lies at present.
Until recently, the computational cost associated with the calculation of risk numbers has been mostly addressed by throwing brute computing capacity at it, that is, buying more and better hardware. It is known that many tier-one banks have farms of several tens of thousand of CPUs and GPUs. Also banks are now leasing cloud computing capability from external vendors. This is, of course, at a considerable cost, which needs to be managed. Obviously, this cannot continue increasing forever without denting the profitability of the business.
Part of the reason why financial institutions have opted for more hardware is down to Moore's second law that states that the computing capacity of transistor chips per dollar of capital expenditure grows exponentially.1 This has certainly been true up to recently. However, this increase in computing capacity was driven by the constant miniaturisation of the basic elements of chips (semiconductor transistors, magnetic memory bits, etc.). Now that we are reaching the 10 nanometer range for semiconductor transistors, the rate of growth stated in Moore's Laws is stalling in commercial computers. This is illustrated by the fact that 10 years ago, the processing capability of a new computer was massively superior compared to the processing capacity of computers only a few years before; at present, it is only marginally better. The reason is that the size of one atom is, roughly, around 0.1 nanometers, and when we decrease the size of transistors below a few tens of nanometers, quantum effects start to appear and temperature becomes a problem. The subject of quantum computing - a most interesting topic - is well outside the scope of this book, but the reality is that, for now and for the foreseeable future, hardware will only be able to offer limited increased computing capacity. As a result, the paradigm has changed from creating more computing power via hardware to developing algorithmic solutions that optimise calculations.
In parallel, there has been a lot of work done by the quantitative analytics community to create algorithmic methods that accelerate the calculations and decrease their hardware need. A notorious example has been the family of Adjoint Algorithmic Differentiation (AAD) solutions in the world of XVA pricing, which in its general version can compute as many XVA sensitivities as needed with the added cost of (roughly) 10 XVA pricing runs. Seen from the perspective of the times when computing one CVA run for a few netting sets was already a challenge, this improvement is remarkable. However, it comes at a considerable price: the implementation effort is most significant. This is particularly the case if one already has a functioning XVA platform and wants to adapt it to AAD. This task can be so daunting that many banks do not consider it a viable option.
This book is based on the belief that the optimal solution to many of the computational challenges in finance lies in the union of algorithmic solutions, and appropriate software implementation of these, run on powerful hardware. The aim of this book is to review how some numerical mathematical methods, when applied thoughtfully, taking into account the specific characteristics of the calculations we want to improve, can create substantial computational enhancements. Indeed, the book is the direct result of the experience the authors have had, over the past few years, while trying to solve difficult calculations (sometimes seemingly impossible) in real-life settings within financial institutions.
The solutions proposed throughout this book apply mainly to existing risk engines within operating financial institutions. Some of these risk engines have been developed over many years by different business units and with different goals in mind. This has produced, in many cases, an amalgamation of risk engines that is suboptimal from an efficiency standpoint. Although starting new engines from scratch may correct the shortcomings of the legacy systems, this requires not only a lot of time and money but also enormous projects in many cases. In fact, a number of banks have reportedly started and stopped the development of global pricing and risk systems from the ground up due to the scale of the job. Quite often, it makes more sense, from a practical perspective, to upgrade existing engines, improving what already exists, using the increasingly demanding business needs and regulatory environment as guidelines, instead of developing a new one. With this in mind, the solutions proposed in this book are highly pragmatic.
Also, we keep in mind that for a solution to be implemented, budgets need to be approved by someone usually high up in the pyramid. Therefore, small(ish) incremental changes with tangible benefits are more likely to succeed than big open ambitious projects. However, to be noted, this does not mean that the solutions put forward in the book cannot be implemented in a system being built from the ground up; in fact, in some cases this would be the optimal approach. All we say is that having the option of incremental changes that are easy to manage is always a bonus to not lose sight of.
One of the common threads in all solutions discussed in the book is that they are grounded on mathematically robust results. Ideally, we would like everything to be based on solid theoretical frameworks. However, as the reader will soon learn, sometimes heuristic rules need to be used in conjunction with mathematical theories. The right combination of mathematical theories and heuristics, partly determined by the context of the problem (for example, the characteristics of the systems being used), is what delivers the most effective outcome. When such heuristic rules are used or discussed, we make the point clear, indicating its range of validity and limitations, so the quantitative analyst can make use of them safely.
Many of the computational problems that banks encounter are the result of having to evaluate a given function a large number of times under (only slightly) different inputs, together with the fact that such functions are costly to compute. Examples of these functions are Over-the-Counter derivative pricing functions, that need to be evaluated from several hundreds to millions of times in risk calculations. From a computational standpoint, these computations tend to be the bottleneck in risk calculations. Our approach is to find a way to take advantage of the specifics in the risk calculation so that a very accurate and fast-to-compute replica of the pricing function can be generated. As a consequence, one computes the same risk metrics in practice, but more efficiently. Also, similar replication methods are applied to other computational challenges for model calibration, for example, leading to significant improvements, too. Furthermore, the techniques presented in this book open the door for a new family of computations that, without them, seem impossible to achieve in many cases, like balance sheet optimisations.
As just said, the solutions discussed in this book are rooted in identifying computationally expensive functions to evaluate - which create computational bottlenecks in calculations - and creating replicas of these problematic functions that can be efficiently computed while at the same time giving essentially the same results.
We start off in Part I with a general overview of Machine Learning techniques. Then we focus on two of the most effective methods to replicate functions: Deep Neural Nets (DNNs) and Chebyshev Tensors (CTs). In mathematical terms, we delve into function approximation because the goal is to create a mathematical object, which comes with a computational architecture, that closely approximates the original function. In our case, we look for techniques that deliver replicas that can be evaluated substantially faster than the function they approximate and that can be calibrated with reasonable...
Dateiformat: ePUBKopierschutz: Adobe-DRM (Digital Rights Management)
Systemvoraussetzungen:
Das Dateiformat ePUB ist sehr gut für Romane und Sachbücher geeignet – also für „fließenden” Text ohne komplexes Layout. Bei E-Readern oder Smartphones passt sich der Zeilen- und Seitenumbruch automatisch den kleinen Displays an. Mit Adobe-DRM wird hier ein „harter” Kopierschutz verwendet. Wenn die notwendigen Voraussetzungen nicht vorliegen, können Sie das E-Book leider nicht öffnen. Daher müssen Sie bereits vor dem Download Ihre Lese-Hardware vorbereiten.Bitte beachten Sie: Wir empfehlen Ihnen unbedingt nach Installation der Lese-Software diese mit Ihrer persönlichen Adobe-ID zu autorisieren!
Weitere Informationen finden Sie in unserer E-Book Hilfe.
