fintech exam
error rate formula
(FP+FN)/(TP+TN+FP+FN)
accuracy formula
(TP + TN)/(TP + TN+ FP + FN)
convolutional neural networks
- good for high dimensional unstructured data - consist of convolutional and pooling layers, as well as fully connected layers -convolution/pooling layers perform automatic feature extraction -fully connected layers then use those features as inputs
hopfield network
backfed input cell
deep belief networks
energy based models app example: recommendation systems
Actor-critic
example of reinforcement learning algorithm
Dynamic programming
example of reinforcement learning algorithm
Policy gradient
example of reinforcement learning algorithm
mean squared erros(MSE)
metrics for unsupervised learning
Denoising autoencoder
noisy input cell- probabilistic hiddent cell - match input output
ai application in asset management
risk management
ai application in asset management
robo-advice
ai application in insurance
robo-advice
volume
scale of data
deep learning applications
sef driving car entertainment visual recognition virtual assistant fraud detection natural language processin g news aggregation detecting development delay in children colourisation of black and white images adding sounds to silent mvies healthcsare personalisations automatic machine transaltion automatic handwriting generaiton demographic and electric predictions automatic game playing language transaltions pixel restoration ohoto desctiptions deep dreaming
ai application in retail and corporate banking
service, credit underwriting and scoring
generative adversarial networks
similar to variational autoencoders, used to learn underlying data distributions and generate new data two networks : generator an discriminator compete in a policeman bandit adversarial game generators create content discriminators clarrify them as fake or authentic purpose: to train the generator to create content as close to authentic as possible
velocity
speed of data generation, procession, and analysis
most common types of neural networks
1. convolutional 2.recurrent NN 3. autoencoders 4.generative adversarial networks
how supervised machine learning works
1. provide the machine learning algorithm categorized or labeled input 2. feed the machine new, unlabeled information to see if it tags new data appropriately. if not, continue refining the algorithm
categorisation of learning algorithms
1. supervised learning 2.unsupervised learning 3. reinforcement learning
Areas in finance most affected by advancements in AI are
1.Asset management 2.Algorithmic trading 3.Blockchain-based applications
lifecycle of AI systems
1.Planning and design, data accumulation and preprocessing, model building 2.Verification and validation 3.Deployment, operation, maintenance
false positive rate formula
FP/(FP+TN)
metrics
Measurements that evaluate algorithms
negative predictive value formula
TN/(TN+FN)
specificity formula
TN/(TN+FP)
Sensitivity formula
TP/(TP+FN)
precision formula
TP/(TP+FP)
reinforcement learning
algorithms (agents) learn a policy that maps states to actions through the optimization of a cumulative return based on trial and error interactions between the agents and their environemnt. input can be labeled or unlabeled
DEEP BELIEF NETWORK (DBN)
backfed input -probabilistic hidden cell- hidden - probabilistic hidden - hiden - probabilistic hidden - match input output cell
Boltzmann Machine
backfed input cell - probabilistic hidden cell
restricted BM
backfed input cell - probabilistic hidden cell
veracity
certainly, reliability
ai application in retail
chat boxes for clients
ai application in insurance
claims management
Supervised Learning types
classification and regression
unsupervised learning types
clustering and dimensionality reduction
ai application in corporate banking
credit loss forecasting, fraud monitoring and detection
ai application in retail and corporate banking
customer service
Q-learning
example of reinforcement learning algorithm
Boosting algorithm
example of supervised learning
decision trees/random forests
example of supervised learning
linear/logistic regression
example of supervised learning
support vector machines
example of supervised learning
Hierarchical aggregating clustering
example of unsupervised learning
K-means
example of unsupervised learning
gaussian mixture models
example of unsupervised learning
Big Data
explosion of quantity of potentially relevant data
Supervised Learning
given data are labeled, ML aims to learn an unknown function that maps features to labels
unsupervised learning
given data are unlabeled, aims to extract similarity patterns
reinforcement learning
goal-directed optimization, software agents interact with their environment aiming to maximize some notion of reward
recurrent neural networks
good for time series input and output data
variety
heterogeneous(images, categorical, real-valued, binary)
Gated Recurrent Unit (GRU)
input - different memory cell- different memory cell - output
kohonen network
input - hidden
Auto encoder
input - hidden - match input output cell
Sparse Autoencoder
input - hidden - match input output cell
SVM (Support Vector Machines)
input - hidden - output
deep residual network
input - hidden - output
NEURAL TURING MACHINE (NTM)
input - hidden - output and blue is match input output cell
deep feed forward
input - hidden cell - output
feed forward neural
input - hidden cell - output
radial basis network
input - hidden cell - output
DCN deep convolutional netowrk
input - kernel - convolutional or pool - hidden - output
Variational Auto Encoder
input - probabilistic hidden cell- match input output cell
recurrent neural networks
input - reccurent cell-recurrent cell - output
liquid state machine
input - spiking hidden cell - output
unsupervised learning
input data are unlabeled: only features
deconvolutional network
input- convolution or pool- kernel - output
Extreme Learning Machine/ELM
input- hidden - output
Generative Adversarial Networks
input- hiden - match input output cell- hiden - match input output cell
deep convolutional inverse graphics network
input- ketnel - convolution - probalistic hidden cell- convolution - karnel - output
LSTM (Long Short Term Memory)
input- memory cell- memory cell - output
ESN echo state network
input- recurrent cell - output
two main types of recurrent neural networks
long short term memory and gated recurrent units
artificial intelligence
machine based systems with varying levels of autonomy that can make predictions, recommendation, or decisions using data
ai application in asset management
management of portfolio strategies
AUC(area under the curve)
metrics for supervised learning
F1 score
metrics for supervised learning
ROC(receiver operating characteristic curve)
metrics for supervised learning
confusion matrix
metrics for supervised learning
precision/recall
metrics for supervised learning
sensitivity
metrics for supervised learning
MAE(Mean Absolute Error)
metrics for unsupervised learning
RMSE
metrics for unsupervised learning
graph neural networks
model the relationship between the nodes in a graph and produce a number representation of it
Deep Learning
neural networks that are implemented over multiple layers of feature extraction and learning
Markov Chain MC
probabilistic hidden cell
trained agent takes proper actions that optimize a cumulative future reward
rainforcement learning metrics
transformers
state of the art method for natural language processing and understanding
machine learning
subset of AI and consists of various algortihms that are able to learn from data
ai application in retail and corporate banking
tailored products
other types of deep NNs
transformer networks graph neural netwroks physics-informed bayesian
autoencoders
used for compessing high dimensional data to low dimensional representations able to learn latent data distribution, used to generate new data
4 Vs: characteristics of big data
volume velocity veracity variety
classification
when labels are binary/categorical. sorting items into categories
regression
when labels are real-valued (dollars, weights, etc_
