we actually predict stock prices with
machine learning investors make educated
guesses by analyzing data will read the
news study the company history industry
trends there are lots of data points
that go into making a prediction the
prevailing theories that stock prices
are totally random and unpredictable a
blindfolded monkey throwing darts at
newspaper’s financial pages could select
a portfolio that would do just as well
as one carefully selected by expert but
that raises the question why did top
firms like Morgan family and Citigroup
higher quantitative analyst to build
predictive models we have this idea of a
trading for being filled with adrenaline
infuse men with highs running around
yelling something is well Colin but
these days are more likely to zero the
machine-learning text first quietly
sitting in front of computer screens in
fact about seventy percent of all
borders on Wall Street are now placed by
software we’re now living in the age of
the algorithm
hello world it’s raj and today we’re
going to build a deep learning model to
predict stock prices records of prices
for traded commodities go back thousands
of years
merchant along popular silk route would
keep records of trade good to try and
predict price trends so that they could
benefit from them and finance the field
of quantitative analysts about 25 years
old and even now it’s still not fully
accepted understood or widely used just
like Google+ it’s a study of how certain
variables correlate with stock price
behavior one of the first attempt this
was made in the seventies by two British
statisticians inbox and Jenkins using
mainframe computers the only historical
data they had access to where prices and
volume they call their model arima and
at the time it was slow and expensive to
run by the eighties things start to get
interesting spreadsheets were invented
to that firms could model company’s
financial performance and automated data
collection became a reality and with
improvements and computing power models
could analyze data much faster it was a
renaissance on Wall Street people were
excited about the possibilities they
started showing up at seminars and
discussing their techniques you should
see what’s going on the bigger firms I
mean I know all the information but all
this quickly die down once people
realize that what works is actually a
very valuable secret right
fuck all about consent the most
successful concept underground in the
past few years we’ve seen a lot of
academic paper published using neural
nets to predict stock prices with
varying degrees of success but until
recently the ability to build these
model has been restricted to academics
who spend their days writing very
complex code now with libraries like
tensorflow anyone can build powerful
predictive model train on massive
datasets so let’s build our own model
using care off with a tensorflow backend
for our training data will be using the
daily closing price of the SMP 500 from
january 2002 August 2016 this is a
series of data points indexed in time
order or a time series
our goal will be to predict the closing
price for any given date after training
we can load are they using a custom load
data function essentially just read our
CSV file into an array of values and
normalizes them rather than being those
values directly to our model normalizing
them improve convergence will use this
equation to normalize each value to
reflect percentage changes from the
starting point so we’ll divide each
price by the initial price and subtract
one when our model later makes
prediction will be normalize the data
using this formula to get a real world
number out of it to build our model will
first initialize it and sequential and
it will be a linear stack of layers then
we’ll add our first layer which is an
lstm layer to what is it
let’s back up for a bit recognize feet
similar to this
you don’t have to think what you did I
already know i sound from opinion it’s
easy to recall the word for word but
could we sing them backwards
no the reason for this is because we
learned these words in a sequence it’s
conditional memory we can access the
word if we act of the word for it
memory matters when we have sequences
are thoughts have persistence but
feed-forward neural Nets don’t accept a
thick sighs vector as input like an
image so we couldn’t use it to say
predict the next frame in a movie
because that would require a sequence of
images vectors as inputs not just one
since the probability of a certain event
happening would depend on what happened
every frame before it we need a way to
allow information to persist and that’s
why we’ll use a recurrent neural net the
current can accept sequences of vectors
of inputs so recall that for
feed-forward neural Nets the hidden
layers wait are based only on the input
data but in a recurrent that the hidden
layer is a combination of the input data
at the current time step and the hidden
layer at a previous time step the hidden
layer is constantly changing as it gets
more inputs and the only way to reach
these hidden States is with the correct
sequence of inputs this is how memory is
incorporated in and we can model this
process mathematically so this hidden
state at a given time that is a function
of the input at that same time step
modified by a weight matrix like the
ones using feed-forward Mets added his
State of the previous time step x its
own hidden state to hidden state matrix
otherwise known as a transition matrix
and because it’s feedback loop is
occurring at every time step in the
series each hidden state has traces of
not only the previous hidden state but
also of all of those that preceded it
that’s why we call it recurrent in a way
we can think of it as copies of the same
network each passing a message to the
so that’s the great thing about
recurrent that they’re able to connect
previous data with the present task but
we still have a problem take a look at
this paragraph it starts off with I hope
senpai will notice me and end with she
is my friend he is my senpai let’s say
we wanted to train a model to predict
this last word given all the other work
we need to contact from the very
beginning of the sequence to know that
this word is probably senpai not
something like buddy or make a regular
recurrent net memories become more
subtle as they ate into the past since
the error signal from later time steps
doesn’t make it far enough back in time
to influence the network at earlier time
steps during backpropagation Joshua NGO
called it the vanishing gradient problem
in one of his most frequently cited
papers piled learning long-term
dependencies with gradient descent is
difficult love the bluntness a popular
solution to this is a modification to
recurring that’s called long short term
memory normally neurons are unit that
apply an activation function like a
sigmoid to a linear combination of their
inputs in an LTM recurrent net we
instead replacing neurons with water
called memory cells each cell has been
implicated in output gate and an
internal state that feeds into itself
across time steps with a constant weight
of one
this eliminates the vanishing gradient
problem since any gradient that flows
into the self recurring units during
backdrop is preserved indefinitely since
errors x 1 still have the same value
HP is an activation function like
signaling during the forward path to
implicate learns when to let activation
packed into the cell and the output
learn to let activation pass out of it
during the backward pass the output get
learns when to let error flow into the
cell and implicate one’s going to let it
flow out of himself through the rest of
the network so despite everything else
in a recurrent that staying the same
doing this more powerful update equation
for our hidden state results in our
network being able to remember long-term
dependencies so for our lstm later we’ll
set our input dimension 21 and say we
want 50 units in this layer setting
return sequences to true means this
layers output is always set into the
next layer all its activations can be
seen as a sequence of predictions the
first layer has made from the input
sequence will add twenty percent drop
out to this layer then initialize our
second layer as another lstm with 100
units and set return sequence to fall on
it since its output is only fed to the
next layer at the end of the sequence it
doesn’t help put a prediction for the
sequence instead a prediction vector for
the whole input sequence will use the
linear dense layer to aggregate the data
from the prediction dr. into one single
value then we can compile our model
using a popular loss function called
mean squared error and use gradient
our optimizer labeled rms prop will
train our model with the function then
we can test it to see what it predicts
for the next 50 steps at several points
in our graph and visualize it using that
pot life it seems that for a lot of the
price movement especially the big one
there is quite a correlation between our
models prediction and the actual data so
time to make some money and place and he
is what will our model be able to
correctly predict the closing price one
hundred percent of the time he’ll to the
no it’s an analytical tool to help us
make educated guesses about the
direction of the market that is slightly
better than random so to break it down
recurrent can model sequential data
since at each time step the hidden state
is affected by the input and the
previous in state solution to the
vanishing gradient problem for recurrent
net is to use long short term memory
cells to remember long-term dependencies
and we can use lstm networks to make
predictions for time series data easily
you can carry off and tensorflow the
winner of the coating talent in the last
video is a shabbat you shall use
transfer learning to create a classifier
cats and dogs he chose the layer from a
pre-training tentacle model and build
his own custom convolutional net on top
of it to make training much factor
wizard of the week and the runner-up is
GS shootin seeds
i loved how he added a command-line
interface for users to input their
images the coding challenge for this
video is to use three different inputs
instead of just one to train your lstm
networks to predict the price of google
stock detailed in the readme poster
gambling in the comments in all about
the winner in a week
please subscribe for more videos like
this and for now i’m going to count my
of layers so thanks for watching