book

Machine Learning for Finance

by James Le, Jannes Klaas

May 2019

Intermediate to advanced

456 pages

11h 38m

English

Packt Publishing

Read now

Unlock full access

Why subscribe?
About the author
Who this book is for
Download the example code filesDownload the color imagesConventions used
Reviews

Our journey in this book
The unreasonable effectiveness of dataAll models are wrong
Running notebooks locallyInstalling TensorFlowInstalling KerasUsing data locally
Python version of our logistic regressor
Gradient descentBackpropagationParameter updatesPutting it all together
Importing KerasA two-layer model in KerasStacking layersCompiling the modelTraining the modelKeras and TensorFlow
The data
Making predictions using the heuristic modelThe F1 scoreEvaluating with a confusion matrix
A feature from intuition – fraudsters don't sleepExpert insight – transfer, then cash outStatistical quirks – errors in balances
One-hot encodingEntity embeddingsTokenizing categoriesCreating input modelsTraining the model
Extracting the targetCreating a test setCreating a validation setOversampling the training dataBuilding the modelCreating a simple baselineBuilding more complex models
A simple decision treeA random forestXGBoost
Convolutional Neural NetworksFilters on MNISTAdding a second filter
Conv2DKernel sizeStride sizePaddingInput shapeSimplified Conv2D notationReLU activationMaxPooling2DFlattenDenseTraining MNISTThe modelLoading the dataCompiling and training
MomentumThe Adam optimizerRegularizationL2 regularizationL1 regularizationRegularization in KerasDropoutBatchnorm
Modifying VGG-16Random image augmentationAugmentation with ImageDataGenerator
Facial recognitionBounding box prediction
Visualization and preparation in pandasAggregate global feature statisticsExamining the sample time series
Different kinds of stationarityWhy stationarity mattersMaking a time series stationaryWhen to ignore stationarity issues
Data preparationWeekdays
The carry
An introductory guide to spaCy
Fine-tuning the NER
Adding custom functions to matchersAdding the matcher to the pipelineCombining rule-based and learning-based systems
Using Python's regex moduleRegex in pandasWhen to use regexes and when not to
Sanitizing charactersLemmatizationPreparing the targetPreparing the training and test sets
TF-IDF
Preprocessing for training with word vectorsLoading pretrained word vectorsTime series models with word vectors
Seq2seq architecture overviewThe dataEncoding charactersCreating inference modelsMaking translations
Understanding autoencodersAutoencoder for MNISTAutoencoder for credit cards
MNIST exampleUsing the Lambda layerKullback–Leibler divergenceCreating a custom lossUsing a VAE to generate dataVAEs for an end-to-end fraud detection system
A MNIST GANUnderstanding GAN latent vectorsGAN training tricks
Using labeling budgets efficientlyLeveraging machines for human labelingPseudo labeling for unlabeled dataUsing generative models
Catch – a quick guide to reinforcement learningQ-learning turns RL into supervised learningDefining the Q-learning modelTraining to play Catch
The Bellman equation in economics
Learning to balanceLearning to trade
Designing good reward functionsCareful, manual reward shapingInverse reinforcement learningLearning from human preferencesRobust RL
Multi-agent RLLearning how to learnUnderstanding the brain through RL
Debugging dataHow to find out whether your data is up to the taskWhat to do if you don't have enough dataUnit testing dataKeeping data private and complying with regulationsPreparing the data for trainingUnderstanding which inputs led to which predictions
Hyperparameter search with HyperasEfficient learning rate searchLearning rate schedulingMonitoring training with TensorBoardExploding and vanishing gradients
Launching fastUnderstanding and monitoring metricsUnderstanding where your data comes from
Using the right hardware for your problemMaking use of distributed training with TF estimatorsUsing optimized layers such as CuDNNLSTMOptimizing your pipelineSpeeding up your code with CythonCaching frequent requests
Sources of unfairness in machine learning
Obtaining causal modelsInstrument variablesNon-linear causal models
Complex systems are intrinsically hazardous systemsCatastrophes are caused by multiple failuresComplex systems run in degraded modeHuman operators both cause and prevent accidentsAccident-free operation requires experience with failure
What is the goal of the model developers?Is the data biased?Are errors biased?How is feedback incorporated?Can the model be interpreted?What happens to models after deployment?
An intuitive guide to Bayesian inferenceFlat prior<50% priorPrior and posteriorMarkov Chain Monte CarloMetropolis-Hastings MCMCFrom probabilistic programming to deep probabilistic programming
General data analysisSound science in machine learningGeneral machine learningGeneral deep learningReinforcement learningBayesian machine learning
Leave a review - let other readers know what you think

Content preview from Machine Learning for Finance

Evolutionary strategies and genetic algorithms

Recently, a decades-old optimization algorithm for reinforcement learning algorithms has come back into fashion. Evolutionary strategies (ES) are much simpler than Q-learning or A2C.

Instead of training one model through backpropagation, in ES we create a population of models by adding random noise to the weights of the original model. We then let each model run in the environment and evaluate its performance. The new model is the performance-weighted average of all the models.

In the following diagram, you can see a visualization of how evolution strategies work: