Chapters will be uploaded as Markdown Files (can be converted to word with Pandoc) or Jupyter Notebooks for chapters that contain lot's of code (can also be converted).

The target reader of this book is a professional working in the financial industry (a domain expert, see Mission). This includes everything from consumer facing banks, to hedge-funds to supporting services such as auditors and accountants. The book assumes basic knowledge of linear algebra and calculus, as well as some basic knowledge of Python. As a rule of thumb, readers who have completed the first two free Python courses by DataCamp should be fine. The book also assumes basic knowledge of micro economics and a rough overview of the financial industry.

The target reader has noticed that the industry is being transformed by machine learning. The reader started learning a bit of Python earlier and has maybe even dabbeled in some other machine learning courses but is unsure of how ML could have an impact in the industry and which skills could be especially valuable. The reader buys this book to get some working knowledge of the most important technologies and find out where to get deeper into.

To better understand the mission and advantage of this book it helps to understand how ML goes from research to value at scale. In terms of business strategy, ML is a horizontal. It is part of multiple value chains across industries. Research so far has focused on strengthening this horizontal and make ML useful for many applications. However, to extract maximum value, ML must be deeply integrated into the different value chains. This will require adopted methods and approaches for each value chain. The task to adopt general ML techniques falls to ML experts and domain experts alike. Domain experts will have to use their knowledge of their domain (in this case the different parts of the financial industry) and knowledge of ML to create custom ML systems.

This book teaches deep learning techniques that are useful in the financial industry in a way that is understandable for domain experts. It uses more of the terminology and tools known to industry professionals and makes use of concepts that are known to financial professionals.

The book takes readers on the journey from a high level understanding of what deep neural networks are, why they work and their limitations to how to practically implement state of the art techniques.

After reading this book readers will understand:

• Neural networks as function approximators
• The gradient descent optimization algorithm
• The value of predictive models to business
• Working with structured data
• Working with image Data
• Working with time series
• Natural language processing
• GANs and other generative models
• Reinforcement learning
• Practicalities of model deployment
• Dealing with ethical and legal concerns around deep learning

1. • A Neural Net from Scratch

2. • Structured Data: From Telemarketing to Fraud Prevention
3. • Deep Learning for Visual Data: From Customer Cards to Searching Mines from Space
4. • From Engines to Stocks: Time series data
5. • Software Robots Reading Reports for you: Natural Language Processing
6. • Generative models for Financial Information Extraction
7. • Reinforcement learning for the Markets

8. • Debugging Neural Nets
9. • Combating Bias in Financial Models

Description: This chapter introduces neural networks as function approximators and gradient descent as a way to find an optimal function. It discusses the flaws of gradient descent and why it is useful in many applications. It concludes with an implementation of a neural network and gradient descent in Python and Excel.

Level: This is a slightly more advanced and complex chapter.

Topics Covered:

• Neural nets as function approximators
• Gradient descent optimization
• Feed forward
• Back propagation

Skills learned: Understanding what neural networks are, how and why they work on the inside.

Description: This chapter introduces the Keras sequential model, and how to use neural networks with structured data.

Level: This is an easy chapter

Topics Covered:

• The Keras sequential model
• Dense layers
• Activation layers
• Data preparation for neural networks
• model.fit()
• Some different optimizers
• Regularization
• Entity embeddings

Skills learned: Build and train simple models with Keras.

Description: This chapter introduces convolutional neural networks and how to work with them in Keras.

Level: This is an average chapter

Topics Covered:

• What are convnets and why should we use them
• Conv2D
• MaxPool2D
• Data generators
• Working with pre-trained models
• Batchnorm
• Object detection & some other advanced architectures

Skills learned: Build and train SOTA image classifiers.

Description: This chapter introduces time series models.

Level: This is a slightly advanced chapter

Topics Covered:

• RNNs
• LSTMS
• Recurrent dropout
• Conv1D
• MaxPool1D
• Sequence classification
• Seq2Seq

Skills learned: Build and train models to forecast and classify from time series.

Description: This chapter introduces natural language processing.

Level: This is an average chapter

Topics Covered:

• Word / Character / N-gram embeddings
• Pre-trained word embeddings: Word2Vec, GloVe
• Preparing text data
• Text classification & Sentiment analysis
• Seq2Seq translation
• Seq2Seq summarization
• Question answering (Note: Only if I find an easy demo implementation somewhere that does not blow the difficulty level)

Skills learned: Build a classification & Seq2Seq NLP systems systems

Description: This chapter shows how to use generative models to extract useful information.

Level: This is an advanced chapter

Topics Covered:

• Autoencoders for unsupervised learning
• Pix2Pix
• Pix2Code
• GANs for semi supervised learning

Skills learned: Build an auto encoder! Translate satellite images to street maps. Generate Website code from screenshots. Create artificial training data. Hussah generative models are fun.

Description: This chapter introduces reinforcement learning and shows some potential paths to 'AI'

Level: This is an advanced chapter

Topics Covered:

• Everything is a markov model
• Q-Learning
• Policy gradients
• Neuro-evolution
• On the importance of good simulators

Skills learned: Build systems that: Play pong. Play atari games. Walk shake and wiggle. Invent trading strategies (and why that might not work)

Description: This chapter shows some practicalities of deployment and debugging

Level: This is an easy chapter

Topics Covered:

• Fighting over / underfit
• Vanishing gradients
• Exploding gradients
• Tensorboard
• Tips on training big models.
• Serving models
• Monitoring model performance

Skills learned: Reduce over/underfit, monitor training with Tensorboard, create custom Keras callbacks, serve Keras with Flask, monitor models and early warning systems.

Description: This chapter discusses ideas about making ML models fair and accountable

Level: This is an easy chapter

Topics Covered:

• Why models amplify biases in data
• Why 'de biasing' data is harder than you think
• Some examples of biased models
• FAT ML research
• Research on black box reasoning
• Why secret models are dangerous but common + alternative strategic approaches.
• Legal situation of black box decision making (EU / US)

Skills learned: Thinking about discrimination issues before building models

Jannes is the lead developer of the Bletchley Bootcamp (see ai-bootcamp.org), a course teaching deep learning to business and economics majors.

1 A Neural Net from Scratch On or before 23/02/2018

2 Structured Data: From Telemarketing to Fraud Prevention On or before 07/03/2018

3 Deep Learning for Visual Data: From Customer Cards to Searching Mines from Space On or before 26/03/2018

4 From Engines to Stocks: Time series data On or before 07/04/2018

5 Software Robots Reading Reports for you: Natural Language Processing On or before 19/04/2018

6 Generative models for Financial Information Extraction On or before 31/04/2018

7 Reinforcement learning for the Markets On or before 12/05/2018

8 Debugging Neural Nets On or before 24/05/2018

9 Combating Bias in Financial Models On or before 05/06/2018

File an issue

Fork, edit, send a pull request. We don't even need separate branches.