Friday, 8 April 2022

Simple Machine Learning Models on OrderBook/PositionBook Features

This post is about using OrderBook/PositionBook features as input to simple machine learning models after previous investigation into the relevance of such features. 

Due to the amount of training data available I decided to look only at a linear model and small neural networks (NN) with a single hidden layer with up to 6 hidden neurons. This choice was motivated by an academic paper I read online about linear models which stated that, as a lower bound, one should have at least 10 training examples for each parameter to be estimated. Other online reading about order flow imbalance (OFI) suggested there is a linear relationship between OFI and price movement. Use of limited size NNs would allow a small amount of non linearity in the relationship. For this investigation I used the Netlab toolbox and Octave. A plot of the learning curves of the classification models tested is shown below. The targets were binary 1/0 for price increases/decreases.

The blue lines show the average training error (y axis) and the red lines show the same average error metric on the held out cross validation data set for each tested model. The thickness of the lines represents the number of neurons in the single hidden layer of the NNs (the thicker the lines, the higher the number of hidden neurons). The horizontal green line shows the error of a generalized linear model (GLM) trained using iteratively reweighted least squares. It can be seen that NN models with 1 and 2 hidden neurons slightly outperform the GLM, with the 2 neuron model having the edge over the 1 neuron model. NN models with 3 or more hidden neurons over fit and underperform the GLM. The NN models were trained using Netlab's functions for Bayesian regularization over the parameters.

Looking at these results it would seem that a 2 neuron NN would be the best choice; however the error differences between the 1 and 2 neuron NNs and GLM are small enough to anticipate that the final classifications (with a basic greater/less than a 0.5 logistic threshold value for long/short) would perhaps be almost identical. 

Investigations into this will be the subject of my next post. 

The code box below gives the working Octave code for the above.

## load data
##training_data = dlmread( 'raw_netlab_training_features' ) ;
##cv_data = dlmread( 'raw_netlab_cv_features' ) ;
training_data = dlmread( 'netlab_training_features_svd' ) ;
cv_data = dlmread( 'netlab_cv_features_svd' ) ;
training_targets = dlmread( 'netlab_training_targets' ) ;
cv_targets = dlmread( 'netlab_cv_targets' ) ;

kk_loop_record = zeros( 30 , 7 ) ;

for kk = 1 : 30
## first train a glm model as a base comparison
input_dim = size( training_data , 2 ) ; ## Number of inputs.

net_lin = glm( input_dim , 1 , 'logistic' ) ; ## Create a generalized linear model structure.
options = foptions ; ## Sets default parameters for optimisation routines, for compatibility with MATLAB's foptions()
options(1) = 1 ;  ## change default value
##	OPTIONS(1) is set to 1 to display error values during training. If
##	OPTIONS(1) is set to 0, then only warning messages are displayed.  If
##	OPTIONS(1) is -1, then nothing is displayed.
options(14) = 5 ; ## change default value
##	OPTIONS(14) is the maximum number of iterations for the IRLS
##	algorithm;  default 100.
net_lin = glmtrain( net_lin , options , training_data , training_targets ) ;

## test on cv_data
glm_out = glmfwd( net_lin , cv_data ) ;
## cross-entrophy loss
glm_out_loss = -mean( cv_targets .* log( glm_out )  .+ ( 1 .- cv_targets ) .* log( 1 .- glm_out ) ) ;

kk_loop_record( kk , 7 ) = glm_out_loss ;

## now train an mlp
## Set up vector of options for the optimiser.
nouter = 30 ; ## Number of outer loops.
ninner = 2 ;	## Number of innter loops.
options = foptions ; ## Default options vector.
options( 1 ) = 1 ;	## This provides display of error values.
options( 2 ) = 1.0e-5 ; ## Absolute precision for weights.
options( 3 ) = 1.0e-5 ; ## Precision for objective function.
options( 14 ) = 100 ; ## Number of training cycles in inner loop.

training_learning_curve = zeros( nouter , 6 ) ; 
cv_learning_curve = zeros( nouter , 6 ) ;

for jj = 1 : 6

## Set up network parameters.
nin = size( training_data , 2 ) ; ## Number of inputs.
nhidden = jj ;	## Number of hidden units.
nout = 1 ; ## Number of outputs.
alpha = 0.01 ; ## Initial prior hyperparameter.
aw1 = 0.01 ;
ab1 = 0.01 ;
aw2 = 0.01 ;
ab2 = 0.01 ;

## Create and initialize network weight vector.
prior = mlpprior(nin , nhidden , nout , aw1 , ab1 , aw2 , ab2 ) ;
net = mlp( nin , nhidden , nout , 'logistic' , prior ) ;

## Train using scaled conjugate gradients, re-estimating alpha and beta.
for ii = 1 : nouter
  ## train net
  net = netopt( net , options , training_data , training_targets , 'scg' ) ;
  train_out = mlpfwd( net , training_data ) ;
  ## get train error
  ## mse
  ##training_learning_curve( ii ) = mean( ( training_targets .- train_out ).^2 ) ;
  ## cross entropy loss
  training_learning_curve( ii , jj ) = -mean( training_targets .* log( train_out )  .+ ( 1 .- training_targets ) .* log( 1 .- train_out ) ) ; 

  cv_out = mlpfwd( net , cv_data ) ;
  ## get cv error
  ## mse
  ##cv_learning_curve( ii ) = mean( ( cv_targets .- cv_out ).^2 ) ;
  ## cross entropy loss
  cv_learning_curve( ii , jj ) = -mean( cv_targets .* log( cv_out )  .+ ( 1 .- cv_targets ) .* log( 1 .- cv_out ) ) ; 
  ## now update hyperparameters based on evidence
  [ net , gamma ] = evidence( net , training_data , training_targets , ninner ) ;
##  fprintf( 1 , '\nRe-estimation cycle ##d:\n' , ii ) ;
##  disp( [ '  alpha = ' , num2str( net.alpha' ) ] ) ;
##  fprintf( 1 , '  gamma =  %8.5f\n\n' , gamma ) ;
##  disp(' ')
##  disp('Press any key to continue.')
endfor ## ii loop

endfor ## jj loop

kk_loop_record( kk , 1 : 6 ) = cv_learning_curve( end , : ) ;

endfor ## kk loop

plot( training_learning_curve(:,1) , 'b' , 'linewidth' , 1 , cv_learning_curve(:,1) , 'r' , 'linewidth' , 1 , ...
training_learning_curve(:,2) , 'b' , 'linewidth' , 2 , cv_learning_curve(:,2) , 'r' , 'linewidth' , 2 , ...
training_learning_curve(:,3) , 'b' , 'linewidth' , 3 , cv_learning_curve(:,3) , 'r' , 'linewidth' , 3 , ...
training_learning_curve(:,4) , 'b' , 'linewidth' , 4 , cv_learning_curve(:,4) , 'r' , 'linewidth' , 4 , ...
training_learning_curve(:,5) , 'b' , 'linewidth' , 5 , cv_learning_curve(:,5) , 'r' , 'linewidth' , 5 , ...
training_learning_curve(:,6) , 'b' , 'linewidth' , 6 , cv_learning_curve(:,6) , 'r' , 'linewidth' , 6 , ...
ones( size( training_learning_curve , 1 ) , 1 ).*glm_out_loss , 'g' , 'linewidth', 2 ) ;

##  >> mean(kk_loop_record)
##  ans =
##     0.6928   0.6927   0.7261   0.7509   0.7821   0.8112   0.6990

##  >> std(kk_loop_record)
##  ans =
##     8.5241e-06   7.2869e-06   1.2999e-02   1.5285e-02   2.5769e-02   2.6844e-02   2.2584e-16

Friday, 25 March 2022

OrderBook and PositionBook Features

In my previous post I talked about how I planned to use constrained optimization to create features from Oanda's OrderBook and PositionBook data, which can be downloaded via their API. In addition to this I have also created a set of features based on the idea of Order Flow Imbalance (OFI), a nice exposition of which is given in this blog post along with a numerical example of how to calculate OFI. Of course Oanda's OrderBook/PositionBook data is not exactly the same as a conventional limit order book, but I thought they are similar enough to investigate using OFI on them. The result of these investigations is shown in the animated GIF below.

This shows the output from using the R Boruta package to check for the feature relevance of OFI levels to a depth of 20 of both the OrderBook and PositionBook to classify the sign of the log return of price over the periods detailed below following an OrderBook/PositionBook update (the granularity at which the OrderBook/PositionBook data can be updated is 20 minutes):

  • 20 minutes
  • 40 minutes
  • 60 minutes
  • the 20 minutes starting 20 minutes in the future
  • the 20 minutes starting 40 minutes in the future
for both the OrderBook and PositionBook, giving a total of 10 separate images/results in the above GIF.
Observant readers may notice that in the GIF there are 42 features being checked, but only an OFI depth of 20. The reason for this is that the data contain information about buys/sell orders and long/short positions both above and below the current price, so what I did was calculate OFI for:
  • buy orders above price vs sell orders below price
  • sell orders above price vs buy orders below price
  • long positions above price vs short positions below price
  • short positions above price vs long positions below price 
As can be seen, almost all features are deemed to be relevant with the exception of 3 OFI levels rejected (red candles) and 2 deemed tentative (yellow candles).

It is my intention to use these features in a machine learning model to classify the probability of future market direction over the time frames mentioned above. 

More in due course.

Tuesday, 15 February 2022

A Possible, New Positionbook Indicator?

In my previous post I ended with saying that I would post about some sort of "sentiment indicator" if, and only if, I had something positive to say about my progress on this work. This post is the first on this subject.

The indicator I'm working on is based on the open position ratios data that is available via the Oanda api. For the uninitiated, this data gives the percentage of traders holding long and short positions, and at what price levels, in 14 selected forex pairs and also gold and silver. The data is updated every 20 minutes. I have long felt that there must be some value hidden in this data but the problem is how to extract it.

What I've done is take the percentage values from the (usually) hundreds of separate price levels and sum and normalise them over three defined ranges - levels above/below the high/low of each 20 minute period and the level(s) that span the price range of this period. This is done separately for long and short positions to give a total of 6 percentage figures that sum to 100%. Conceptually, this can be thought of as attaching to the open and close of a 20 minute OHLC bar the 6 percentage position values that were in force at the open and close respectively. The problem is to try and infer the actual, net changes in positions that have taken place over the time period this 20 minute bar was forming. In this way I am trying, if you like, to create a sort of  "skin in the game" indicator as opposed to an indicator derived from order book data, which could be said to be based on traders' current (changeable) intentions as expressed by their open orders and which are subject to shenanigans such as spoofing.

The methodology I've decided on to realise the above is constrained optimization using Octave's fmincon function. The objective function is simply:

    denom = X' * old_pb_net_pos ;
    J = mean( ( new_pb_net_pos .- ( ( X .* old_pb_net_pos ) ./ denom ) ).^2 ) ;

for a multiplicative position value change model where:

  • X is a vector of constants that are to be optimised
  • old_pb_net_pos is a vector of the 6 percentage values at the open
  • new_pb_net_pos is a vector of the 6 percentage values at the close

This is a constrained model because percentage position values at price levels outside the bar range cannot actually increase as a result of trades that take place within the bar range, so the X values for these levels are necessarily constrained to a maximum value of 1 (implying no real, absolute change at these levels). Similarly, all X values must be greater than zero (a zero value would imply a mass exit of all positions at this level, which never actually happens).

The net result of the above is an optimised X vector consisting of multiplicative constants that are multiplied with old_pb_net_pos to achieve new_pb_net_pos according to the logic exemplified in the above objective function. It is these optimised X values from which the underlying, real changes in positions will be inferred and features created. More on this in my next post.


Tuesday, 4 January 2022

Matrix Profile and Weakly Labelled Data - 2nd and Final Update

It has been over three months since my last post, which was intended to be the first in a series of posts on the subject of the title of this post. However, it turned out that the results of my work were underwhelming and so I decided to stop flogging a dead horse and move onto other things. I still have some ideas for using Matrix Profile, but not for the above. These ideas may be the subject of a future blog post.

I subsequently looked at plotting order levels using the data that is available via the Oanda API and I have come up with Octave code to render plots such as this:

where the brighter yellow stripes show ranges where there is an accumulation of sell/buy orders above/below price. These can be interpreted as support/resistance areas. It is normally my practice to post my Octave code, but the code for this plot is quite idiosyncratic and depends very much on the way I have chosen to store the underlying data downloaded from Oanda. As such, I don't think it would be helpful to readers and so I am not posting the code. That said, if there is actually a demand I am more than happy to make it available in a future blog post.

Having done this, it seemed natural to extend it to Open Position Ratios which are also available via the Oanda API. Plotting these levels renders plots that are similar to the plot shown above, but show levels where open long/short positions instead of open orders are accumulated. Although such plots are visually informative, I prefer something more objective, and so for the last few weeks I have been working on using the open position ratios data to construct some sort of sentiment indicator that hopefully could give a heads up to future price movement direction. This is still very much a work in progress which I shall post about if there are noteworthy results.

More in due course.

Friday, 17 September 2021

Matrix Profile and Weakly Labelled Data - Update 1

This is the first post in a short series detailing my recent work following on from my previous post. This post will be about some problems I have had and how I partially solved them.

The main problem was simply the speed at which the code (available from the companion website) seems to run. The first stage Matrix Profile code runs in a few seconds, the second, individual evaluation stage in no more than a few minutes, but the third stage, greedy search, which uses Golden Section Search over the pattern candidates, can take many, many hours. My approach to this was simply to optimise the code to the best of my ability. My optimisations, all in the compute_f_meas.m function, are shown in the following code boxes. This while loop

i = 1;
while true
 if i >= length(anno_st)
first_part = anno_st(1:i);
second_part = anno_st(i+1:end);
bad_st = abs(second_part - anno_st(i)) < sub_len;
second_part = second_part(~bad_st);
anno_st = [first_part; second_part;];
i = i + 1;
is replaced by this .oct compiled version of the same while loop

DEFUN_DLD ( stds_f_meas_while_loop_replace, args, nargout,
"-*- texinfo -*-\n\
@deftypefn {Function File} {} stds_f_meas_while_loop_replace (@var{input_vector,sublen})\n\
This function takes an input vector and a scalar sublen\n\
length. The function sets to zero those elements in the\n\
input vector that are closer to the preceeding value than\n\
sublen. This function replaces a time consuming .m while loop\n\
in the stds compute_f_meas.m function.\n\
@end deftypefn" )

octave_value_list retval_list ;
int nargin = args.length () ;

// check the input arguments
if ( nargin != 2 ) // there must be a vector and a scalar sublen
   error ("Invalid arguments. Inputs are a column vector and a scalar value sublen.") ;
   return retval_list ;

if ( args(0).length () < 2 )
   error ("Invalid 1st argument length. Input is a column vector of length > 1.") ;
   return retval_list ;
if ( args(1).length () > 1 )
   error ("Invalid 2nd argument length. Input is a scalar value for sublen.") ;
   return retval_list ;
// end of input checking  
ColumnVector input = args(0).column_vector_value () ;
double sublen = args(1).double_value () ;
double last_iter ;

// initialise last_iter value
last_iter = input( 0 ) ;
for ( octave_idx_type ii ( 1 ) ; ii < args(0).length () ; ii++ )
      if ( input( ii ) - last_iter >= sublen )
        last_iter = input( ii ) ;
        input( ii ) = 0.0 ;
    } // end for loop
retval_list( 0 ) = input ;

return retval_list ;

} // end of function
and called thus
anno_st = stds_f_meas_while_loop_replace( anno_st , sub_len ) ;
anno_st( anno_st == 0 ) = [] ;
This for loop
is_tp = false(length(anno_st), 1);
for i = 1:length(anno_st)
    if anno_ed(i) > length(label)
        anno_ed(i) = length(label);
    if sum(label(anno_st(i):anno_ed(i))) > 0.8*sub_len
        is_tp(i) = true;
tp_pre = sum(is_tp);
is replaced by use of cellslices.m and cellfun.m thus
label_length = length( label ) ;
anno_ed( anno_ed > label_length ) = label_length ;
cell_slices = cellslices( label , anno_st , anno_ed ) ;
cell_sums = cellfun( @sum , cell_slices ) ;
tp_pre = sum( cell_sums > 0.8 * sub_len ) ;
and a further for loop
is_tp = false(length(pos_st), 1);
for i = 1:length(pos_st)
    if sum(anno(pos_st(i):pos_ed(i))) > 0.8*sub_len
        is_tp(i) = true;
tp_rec = sum(is_tp);
is replaced by
cell_slices = cellslices( anno , pos_st , pos_ed ) ;
cell_sums = cellfun( @sum , cell_slices ) ;
tp_rec = sum( cell_sums > 0.8 * sub_len ) ;

Although the above measurably improves running times, overall the code of the third stage is still sluggish. I have found that the best way to deal with this, on the advice of the original paper's author, is to limit the number of patterns to search for, the "pat_max" variable, to the minimum possible to achieve a satisfactory result. What I mean by this is that if  pat_max = 5 and the result returned also has 5 identified patterns, incrementally increase pat_max until such time that the number of identified patterns is less than pat_max. This does, by necessity, mean running the whole routine a few times, but it is still quicker this way than drastically over estimating pat_max, i.e. choosing a value of say 50 to finally identify maybe only 5/6 patterns.

More in due course.

Saturday, 4 September 2021

"Matrix profile: Using Weakly Labeled Time Series to Predict Outcomes" Paper

Back in May of this year I posted about how I had intended to use Matrix Profile (MP) to somehow cluster the "initial balance" of Market Profile charts with a view to getting a heads up on immediately following price action. Since then, my thinking has evolved due to my learning about the paper "Matrix profile: Using Weakly Labeled Time Series to Predict Outcomes" and its companion website. This very much seems to accomplish the same end I had envisaged with my clustering of initial balances, so I am going to try and use this approach instead.

As a preliminary, I have decided to "weakly label" my time series data using the simple code loop shown below.

for ii = 1 : numel( ix )
y_values = train_data( ix( ii ) + 1 : ix( ii ) + 19 , 1 ) ;
london_session_ret = y_values( end ) - y_values( 1 ) ;

[ max_y , max_ix ] = max( y_values ) ;
max_long_ex = max_y - y_values( 1 ) ;

[ min_y , min_ix ] = min( y_values ) ;
max_short_ex = min_y - y_values( 1 ) ;

if ( london_session_ret > 0 && ( max_long_ex / ( -1 * max_short_ex ) ) >= 3 && max_ix > min_ix )
 labels( ix( ii ) - 11 : ix( ii ) , 1 ) = 1 ; 
elseif ( london_session_ret < 0 && ( max_short_ex / max_long_ex ) <= -3 && max_ix < min_ix )
  labels( ix( ii ) - 11 : ix( ii ) , 1 ) = -1 ;
What this essentially does (for the long side) is ensure that price is higher at the end of y_values than at the beginning and there is a reward/risk opportunity of at least 3:1 for at least 1 trade during the period covered by the time range of y_values (either the London a.m. session or the combined New York a.m./London p.m. session) following a 7a.m. to 8.50a.m. (local time) formation of an opening Market profile/initial balance and the maximum adverse excursion occurs before the maximum favourable excursion. A typical chart on the long side looks like this.
This would have the "weak" label for a long trade, and the label would be applied to the Market Profile data that immediately precedes this price action. On the other side, a short labelled chart typically looks like this.
As can be seen, trading "against the label" offers few opportunities for profitable entries/exits. My hope is that a "dictionary" of long/short biased Market Profile patterns can be discovered using the ideas/code in the links above. For completeness, the following chart is typical of price action which does not meet the looped code bias for either long or short.

It is easy to envisage trading this type of price action by fading moves that go outside the "value area" of a Market Profile chart.

More in due course.

Friday, 27 August 2021

Another Iterative Improvement of my Volume/Market Profile Charts

Below is a screenshot of this new chart version, of today's (Friday's) price action at a 10 minute bar scale:

Just by looking at the chart it might not be obvious to readers what has changed, so the changes are detailed below.

The first change is in how the volume profile (the horizontal histogram on the left) is calculated. The "old" version of the chart calculates the profile by assuming the "model" that tick volume for each 10 minute bar is normally distributed across the high/low range of the bar, and then the profile histogram is the accumulation of these individual, 10 minute, normally distributed "mini profiles." A more complete description of this is given in my Market Profile Chart in Octave blog post, with code.

The new approach is more data centric rather than model based. Every 10 minutes, instead of downloading the 10 minute OHLC and tick volume, the last 10 minutes worth of 5 second OHLC and tick volume is downloaded. The whole tick volume of each 5 second period is assigned to a price level equivalent to the Typical price (rounded to the nearest pip) of said 5 second period, and the volume profile is then the accumulation of these volume ticks per price level. I think this is a much more accurate reflection of the price levels at which tick volume actually occurred compared to the old, model based charts. This second screenshot is of the old chart over the exact same price data as the first, improved version of the chart.

It can be seen that the two volume profile histograms of the respective charts differ from each other in terms of their overall shape and the number and price levels of peaks (Points of Control) and troughs (Low Volume Nodes).

The second change in the new chart is in how the background heatmap is plotted. The heatmap is a different presentation of the volume profile whereby higher volume price levels are shown by the brighter yellow colours. The old chart only displays the heatmap associated with the latest calculated volume profile histogram, which is projected back in time. This is, of course, a form of lookahead bias when plotting past prices over the latest heatmap. The new chart solves this by plotting a "rolling" version of the heatmap which reflects the volume profile that was in force at the time each 10 minute OHLC candle formed. It is easy to see how the Points of Control and Low Volume Nodes price levels ebb and flow throughout the trading day.

The third change, which naturally followed on from the downloading of 5 second data, is in the plotting of the candlesticks. Rather than having a normal, open to close candlestick body, the candlesticks show the "mini volume profiles" of the tick volume within each bar, plotted via Octave's patch function. The white candlestick wicks indicate the usual high/low range, and the open and close levels are shown by grey and black dots respectively. This is more clearly seen in the zoomed in screenshot below.

I wanted to plot these types of bars because recently I have watched some trading webcasts, which talked about "P", "b" and "D" shaped bar profiles at "areas of interest." The upshot of these webcasts is that, in general, a "P" bar is bullish, a "b" is bearish and a "D" is "in balance" when they intersect an "area of interest" such as Point of Control, Low Volume Node, support and resistance etc. This is supposed to be indicative of future price direction over the immediate short term. With this new version of chart, I shall be in a position to investigate these claims for myself.