I would like to identify the largest possible contiguous subsample of a large data set. My data set consists of roughly 15,000 financial time series of up to 360 periods in length. I have imported the data into MATLAB as a 360 by 15,000 numerical matrix.
This matrix contains a lot of NaNs due to some of the financial data not being available for the entire period. In the illustration, NaN entries are shown in dark blue, and non-NaN entries appear in light blue. It is these light blue non-NaN entries which I would like to ideally combine into an optimal subsample.
I would like to find the largest possible contiguous block of data that is contained in my matrix, while ensuring that my matrix contains a sufficient number of periods.
In a first step I would like to sort my matrix from left to right in descending order by the number of non-NaN entries in each column, that is, I would like to sort by the vector obtained by entering sum(~isnan(data),1).
In a second step I would like to find the sub-array of my data matrix that is at least 72 entries along the first dimension and is otherwise as large as possible, measured by the total number of entries.
What is the best way to implement this?
A big warning (may or may not apply depending on context)
As Oleg mentioned, when an observation is missing from a financial time series, it's often missing for reason: eg. the entity went bankrupt, the entity was delisted, or the instrument did not trade (i.e. illiquid). Constructing a sample without NaNs is likely equivalent to constructing a sample where none of these events occur!
For example, if this were hedge fund return data, selecting a sample without NaNs would exclude funds that blew up and ceased trading. Excluding imploded funds would bias estimates of expected returns upwards and estimates of variance or covariance downwards.
Picking a sample period with the fewest time series with NaNs would also exclude periods like the 2008 financial crisis, which may or may not make sense. Excluding 2008 could lead to an underestimate of how haywire things could get (though including it could lead to overestimate the probability of certain rare events).
Some things to do:
Pick a sample period as long as possible but be aware of the limitations.
Do your best to handle survivorship bias: eg. if NaNs represent delisting events, try to get some kind of delisting return.
You almost certainly will have an unbalanced panel with missing observations, and your algorithm will have to be deal with that.
Another general finance / panel data point, selecting a sample at some time point t and then following it into the future is perfectly ok. But selecting a sample based upon what happens during or after the sample period can be incredibly misleading.
Code that does what you asked:
This should do what you asked and be quite fast. Be aware of the problems though if whether an observation is missing is not random and orthogonal to what you care about.
Inputs are a T by n sized matrix X:
T = 360; % number of time periods (i.e. rows) in X
n = 15000; % number of time series (i.e. columns) in X
T_subsample = 72; % desired length of sample (i.e. rows of newX)
% number of possible starting points for series of length T_subsample
nancount_periods = T - T_subsample + 1;
nancount = zeros(n, nancount_periods, 'int32'); % will hold a count of NaNs
X_isnan = int32(isnan(X));
nancount(:,1) = sum(X_isnan(1:T_subsample, :))'; % 'initialize
% We need to obtain a count of nans in T_subsample sized window for each
% possible time period
j = 1;
for i=T_subsample + 1:T
% One pass: add new period in the window and subtract period no longer in the window
nancount(:,j+1) = nancount(:,j) + X_isnan(i,:)' - X_isnan(j,:)';
j = j + 1;
end
indicator = nancount==0; % indicator of whether starting_period, series
% has no NaNs
% number of nonan series of length T_subsample by starting period
max_subsample_size_by_starting_period = sum(indicator);
max_subsample_size = max(max_subsample_size_by_starting_period);
% find the best starting period
starting_period = find(max_subsample_size_by_starting_period==max_subsample_size, 1);
ending_period = starting_period + T_subsample - 1;
columns_mask = indicator(:,starting_period);
columns = find(columns_mask); %holds the column ids we are using
newX = X(starting_period:ending_period, columns_mask);
Here's an idea,
Assuming you can rearrange the series, calculate the distance (you decide the metric, but if looking at is nan vs not is nan, Hamming is ok).
Now hierarchically cluster the series and rearrange them using either a dendrogram
or http://www.mathworks.com/help/bioinfo/examples/working-with-the-clustergram-function.html
You should probably prune any series that doesn't have a minimum number of non nan values before you start.
First I have only little insight in financial mathematics. I understood it that you want to find the longest continuous chain of non-NaN values for each time series. The time series should be sorted depending on the length of this chain and each time series, not containing a chain above a threshold, discarded. This can be done using
data = rand(360,15e3);
data(abs(data) <= 0.02) = NaN;
%% sort and chop data based on amount of consecutive non-NaN values
binary_data = ~isnan(data);
% find edges, denote their type and calculate the biggest chunk in each
% column
edges = [2*binary_data(1,:)-1; diff(binary_data, 1)];
chunk_size = diff(find(edges));
chunk_size(end+1) = numel(edges)-sum(chunk_size);
[row, ~, id] = find(edges);
num_row_elements = diff(find(row == 1));
num_row_elements(end+1) = numel(chunk_size) - sum(num_row_elements);
%a chunk of NaN has a -1 in id, a chunk of non-NaN a 1
chunks_per_row = mat2cell(chunk_size .* id,num_row_elements,1);
% sort by largest consecutive block of non-NaNs
max_size = cellfun(#max, chunks_per_row);
[max_size_sorted, idx] = sort(max_size, 'descend');
data_sorted = data(:,idx);
% remove all elements that only have block sizes smaller then some number
some_number = 20;
data_sort_chop = data_sorted(:,max_size_sorted >= some_number);
Note that this can be done a lot simpler, if the order of periods within a time series doesn't matter, aka data([1 2 3],id) and data([3 1 2], id) are identical.
What I do not know is, if you want to discard all periods within a time series that don't correspond to the biggest value, get all those chains as individual time series, ...
Feel free to drop a comment if it has to be more specific.
Related
I have this piece of code
N=10^4;
for i = 1:N
[E,X,T] = fffun(); % Stochastic simulation. Returns every time three different vectors (whose length is 10^3).
X_(i,:)=X;
T_(i,:)=T;
GRID=[GRID T];
end
GRID=unique(GRID);
% Second part
for i=1:N
for j=1:(kmax)
f=find(GRID==T_(i,j) | GRID==T_(i,j+1));
s=f(1);
e=f(2)-1;
counter(X_(i,j), s:e)=counter(X_(i,j), s:e)+1;
end
end
The code performs N different simulations of a stochastic process (which consists of 10^3 events, occurring at discrete moments (T vector) that depends on the specific simulation.
Now (second part) I want to know, as a function of time istant, how many simulations are in a particular state (X assumes value between 1 and 10). The idea I had: create a grid vector with all the moments at which something happens in any simulation. Then, looping over the simulations, loop over the timesteps in which something happens and incrementing all the counter indeces that corresponds to this particular slice of time.
However this second part is very heavy (I mean days of processing on a standard quad-core CPU). And it shouldn't.
Are there any ideas (maybe about comparing vectors in a more efficient way) to cut the CPU time?
This is a standalone 'second_part'
N=5000;
counter=zeros(11,length(GRID));
for i=1:N
disp(['Counting sim #' num2str(i)]);
for j=1:(kmax)
f=find(GRID==T_(i,j) | GRID==T_(i,j+1),2);
s=f(1);
e=f(2)-1;
counter(X_(i,j), s:e)=counter(X_(i,j), s:e)+1;
end
end
counter=counter/N;
stop=find(GRID==Tmin);
stop=stop-1;
plot(counter(:,(stop-500):stop)')
with associated dummy data ( filedropper.com/data_38 ). In the real context the matrix has 2x rows and 10x columns.
Here is what I understand:
T_ is a matrix of time steps from N simulations.
X_ is a matrix of simulation state at T_ in those simulations.
so if you do:
[ut,~,ic]= unique(T_(:));
you get ic which is a vector of indices for all unique elements in T_. Then you can write:
counter = accumarray([ic X_(:)],1);
and get counter with no. of rows as your unique timesteps, and no. of columns as the unique states in X_ (which are all, and must be, integers). Now you can say that for each timestep ut(k) the number of time that the simulation was in state m is counter(k,m).
In your data, the only combination of m and k that has a value greater than 1 is (1,1).
Edit:
From the comments below, I understand that you record all state changes, and the time steps when they occur. Then every time a simulation change a state you want to collect all the states from all simulations and count how many states are from each type.
The main problem here is that your time is continuous, so basically each element in T_ is unique, and you have over a million time steps to loop over. Fully vectorizing such a process will need about 80GB of memory which will probably stuck your computer.
So I looked for a combination of vectorizing and looping through the time steps. We start by finding all unique intervals, and preallocating counter:
ut = unique(T_(:));
stt = 11; % no. of states
counter = zeros(stt,numel(ut));r = 1:size(T_,1);
r = 1:size(T_,1); % we will need that also later
Then we loop over all element in ut, and each time look for the relevant timestep in T_ in all simulations in a vectorized way. And finally we use histcounts to count all the states:
for k = 1:numel(ut)
temp = T_<=ut(k); % mark all time steps before ut(k)
s = cumsum(temp,2); % count the columns
col_ind = s(:,end); % fins the column index for each simulation
% convert the coulmns to linear indices:
linind = sub2ind(size(T_),r,col_ind.');
% count the states:
counter(:,k) = histcounts(X_(linind),1:stt+1);
end
This takes about 4 seconds at my computer for 1000 simulations, so it adds to a little more than one hour for the whole process. Not very quick...
You can try also one or two of the tweaks below to squeeze run time a little bit more:
As you can read here, accumarray seems to work faster in small arrays then histcouns. So may want to switch to it.
Also, computing linear indices directly is a quicker method than sub2ind, so you may want to try that.
implementing these suggestions in the loop above, we get:
R = size(T_,1);
r = (1:R).';
for k = 1:K
temp = T_<=ut(k); % mark all time steps before ut(k)
s = cumsum(temp,2); % count the columns
col_ind = s(:,end); % fins the column index for each simulation
% convert the coulmns to linear indices:
linind = R*(col_ind-1)+r;
% count the states:
counter(:,k) = accumarray(X_(linind),1,[stt 1]);
end
In my computer switching to accumarray and or removing sub2ind gain a slight improvement but it was not consistent (using timeit for testing on 100 or 1K elements in ut), so you better test it yourself. However, this still remains very long.
One thing that you may want to consider is trying to discretize your timesteps, so you will have much less unique elements to loop over. In your data about 8% of the time intervals a smaller than 1. If you can assume that this is short enough to be treated as one time step, then you could round your T_ and get only ~12.5K unique elements, which take about a minute to loop over. You can do the same for 0.1 intervals (which are less than 1% of the time intervals), and get 122K elements to loop over, what will take about 8 hours...
Of course, all the timing above are rough estimates using the same algorithm. If you do choose to round the times there may be even better ways to solve this.
Didn't know how to paraphrase the question well.
Function for example:
Data:https://www.dropbox.com/s/wr61qyhhf6ujvny/data.mat?dl=0
In this case how do I calculate that the rest point of this function is ~1? I have access to the vector that makes the plot.
I guess the mean is an approximation but in some cases it can be pretty bad.
Under the assumption that the "rest" point is the steady-state value in your data and the fact that the steady-state value happens the majority of the times in your data, you can simply bin all of the points and use each unique value as a separate bin. The bin with the highest count should correspond to the steady-state value.
You can do this by a combination of histc and unique. Assuming your data is stored in y, do this:
%// Find all unique values in your data
bins = unique(y);
%// Find the total number of occurrences per unique value
counts = histc(y, bins);
%// Figure out which bin has the largest count
[~,max_bin] = max(counts);
%// Figure out the corresponding y value
ss_value = bins(max_bin);
ss_value contains the steady-state value of your data, corresponding to the most occurring output point with the assumptions I laid out above.
A minor caveat with the above approach is that this is not friendly to floating point data whose unique values are generated by floating point values whose decimal values beyond the first few significant digits are different.
Here's an example of your data from point 2300 to 2320:
>> format long g;
>> y(2300:2320)
ans =
0.99995724232555
0.999957488454868
0.999957733165346
0.999957976465197
0.999958218362579
0.999958458865564
0.999958697982251
0.999958935720613
0.999959172088623
0.999959407094224
0.999959640745246
0.999959873049548
0.999960104014889
0.999960333649014
0.999960561959611
0.999960788954326
0.99996101464076
0.999961239026462
0.999961462118947
0.999961683925704
0.999961904454139
Therefore, what I'd recommend is to perhaps round so that the first 5 or so significant digits are maintained.
You can do this to your dataset before you continue:
num_digits = 5;
y_round = round(y*(10^num_digits))/(10^num_digits);
This will first multiply by 10^n where n is the number of digits you desire so that the decimal point is shifted over by n positions. We round this result, then divide by 10^n to bring it back to the scale that it was before. If you do this, for those points that were 0.9999... where there are n decimal places, these will get rounded to 1, and it may help in the above calculations.
However, more recent versions of MATLAB have this functionality already built-in to round, and you can just do this:
num_digits = 5;
y_round = round(y,num_digits);
Minor Note
More recent versions of MATLAB discourage the use of histc and recommend you use histcounts instead. Same function definition and expected inputs and outputs... so just replace histc with histcounts if your MATLAB version can handle it.
Using the above logic, you could also use the median too. If the majority of data is fluctuating around 1, then the median would have a high probability that the steady-state value is chosen... so try this too:
ss_value = median(y_round);
I am attempting to create a model whereby there is a line - represented as a 1D matrix populated with 1's - and points on the line are generated at random. Every time a point is chosen (A), it creates a 'zone of exclusion' (based on an exponential function) such that choosing another point nearby has a much lower probability of occurring.
Two main questions:
(1) What is the best way to generate an exponential such that I can multiply the numbers surrounding the chosen point to create the zone of exclusion? I know of exppdf however i'm not sure if this allows me to create an exponential which terminates at 1, as I need the zone of exclusion to end and the probability to return to 1 eventually.
(2) How can I modify matrix values plus/minus a specific index (including that index)? I got as far as:
x(1:100) = 1; % Creates a 1D-matrix populated with 1's
p = randi([1 100],1,1);
x(p) =
But am not sure how to go about using the randomly generated number to alter values in the matrix.
Any help would be much appreciated,
Anna
Don't worry about exppdf, pick the width you want (how far away from the selected point does the probability return to 1?) and define some simple function that makes a small vector with zero in the middle and 1 at the edges. So here I'm just modifying a section of length 11 centred on p and doing nothing to the rest of x:
x(1:100)=1;
p = randi([1 100],1,1);
% following just scaled
somedist = (abs(-5:5).^2)/25;
% note - this will fail if p is at edges of data, but see below
x(p-5:p+5)=x(p-5:p+5).*somedist;
Then, instead of using randi to pick points you can use datasample which allows for giving weights. In this case your "data" is just the numbers 1:100. However, to make edges easier I'd suggest initialising with a "weight" vector which has zero padding - these sections of x will not be sampled from but stop you from having to make edge checks.
x = zeros([1 110]);
x(6:105)=1;
somedist = (abs(-5:5).^2)/25;
nsamples = 10;
for n = 1:nsamples
p = datasample(1:110,1,'Weights',x);
% if required store chosen p somewhere
x(p-5:p+5)=x(p-5:p+5).*somedist;
end
For an exponential exclusion zone you could do something like:
somedist = exp(abs(-5:5))/exp(5)-exp(0)/exp(5);
It doesn't quite return to 1 but fairly close. Here's the central region of x (ignoring the padding) after two separate runs:
An issue I've come across multiple times is wanting to take two similar data sets and create histograms from them where the bins are identical, so as to easily calculate things like histogram overlap.
You can define the number of bins (obviously) using
[counts, bins] = hist(data,number_of_bins)
But there's not an obvious way (as far as I can see) to make the bin size equal for several different data sets. If remember when I initially looked finding various people who seem to have the same issue, but no good solutions.
The right, easy way
As pointed out by horchler, this can easily be achieved using either histc (which lets you define your bins vector), or vectorizing your histogram input into hist.
The wrong, stupid way
I'm leaving below as a reminder to others that even stupid questions can yield worthwhile answers
I've been using the following approach for a while, so figured it might be useful for others (or, someone can very quickly point out the correct way to do this!).
The general approach relies on the fact that MATLAB's hist function defines an equally spaced number of bins between the largest and smallest value in your sample. So, if you append a start (smallest) and end (largest) value to your various samples which is the min and max for all samples of interest, this forces the histogram range to be equal for all your data sets. You can then truncate the first and last values to recreate your original data.
For example, create the following data set
A = randn(1,2000)+7
B = randn(1,2000)+9
[counts_A, bins_A] = hist(A, 500);
[counts_B, bins_B] = hist(B, 500);
Here for my specific data sets I get the following results
bins_A(1) % 3.8127 (this is also min(A) )
bins_A(500) % 10.3081 (this is also max(A) )
bins_B(1) % 5.6310 (this is also min(B) )
bins_B(500) % 13.0254 (this is also max(B) )
To create equal bins you can simply first define a min and max value which is slightly smaller than both ranges;
topval = max([max(A) max(B)])+0.05;
bottomval = min([min(A) min(B)])-0.05;
The addition/subtraction of 0.05 is based on knowledge of the range of values - you don't want your extra bin to be too far or too close to the actual range. That being said, for this example by using the joint min/max values this code will work irrespective of the A and B values generated.
Now we re-create histogram counts and bins using (note the extra 2 bins are for our new largest and smallest value)
[counts_Ae, bins_Ae] = hist([bottomval, A, topval], 502);
[counts_Be, bins_Be] = hist([bottomval, B, topval], 502);
Finally, you truncate the first and last bin and value entries to recreate your original sample exactly
bins_A = bins_Ae(2:501)
bins_B = bins_Ae(2:501)
counts_A = counts_Ae(2:501)
counts_B = counts_Be(2:501)
Now
bins_A(1) % 3.7655
bins_A(500) % 13.0735
bins_B(1) % 3.7655
bins_B(500) % 13.0735
From this you can easily plot both histograms again
bar([bins_A;bins_B]', [counts_A;counts_B]')
And also plot the histogram overlap with ease
bar(bins_A,(counts_A+counts_B)-(abs(counts_A-counts_B)))
just wondering if anyone has any ideas about an issue I'm having.
I have a fair amount of data that needs to be displayed on one graph. Two theoretical lines that are bold and solid are displayed on top, then 10 experimental data sets that converge to these lines are graphed, each using a different identifier (eg the + or o or a square etc). These graphs are on a log scale that goes up to 1e6. The first few decades of the graph (< 1e3) look fine, but as all the datasets converge (> 1e3) it's really difficult to see what data is what.
There's over 1000 data points points per decade which I can prune linearly to an extent, but if I do this too much the lower end of the graph will suffer in resolution.
What I'd like to do is prune logarithmically, strongest at the high end, working back to 0. My question is: how can I get a logarithmically scaled index vector rather than a linear one?
My initial assumption was that as my data is lenear I could just use a linear index to prune, which lead to something like this (but for all decades):
//%grab indicies per decade
ind12 = find(y >= 1e1 & y <= 1e2);
indlow = find(y < 1e2);
indhigh = find(y > 1e4);
ind23 = find(y >+ 1e2 & y <= 1e3);
ind34 = find(y >+ 1e3 & y <= 1e4);
//%We want ind12 indexes in this decade, find spacing
tot23 = round(length(ind23)/length(ind12));
tot34 = round(length(ind34)/length(ind12));
//%grab ones to keep
ind23keep = ind23(1):tot23:ind23(end);
ind34keep = ind34(1):tot34:ind34(end);
indnew = [indlow' ind23keep ind34keep indhigh'];
loglog(x(indnew), y(indnew));
But this causes the prune to behave in a jumpy fashion obviously. Each decade has the number of points that I'd like, but as it's a linear distribution, the points tend to be clumped at the high end of the decade on the log scale.
Any ideas on how I can do this?
I think the easiest way to do this would be to use the LOGSPACE function to generate a set of indices into your data. For example, to create a set of 100 points logarithmically spaced from 1 to N (the number of points in your data), you can try the following:
indnew = round(logspace(0,log10(N),100)); %# Create the log-spaced index
indnew = unique(indnew); %# Remove duplicate indices
loglog(x(indnew),y(indnew)); %# Plot the indexed data
Creating a logarithmically-spaced index like this will result in fewer values being chosen from the end of the vector relative to the start, thus pruning values more severely towards the end of the vector and improving the appearance of the log plot. It would therefore be most effective with vectors that are sorted in ascending order.
The way I understand the problem is that your x-values are linearly spaced, so that if you plot them logarithmically, there are way more data points in 'higher' decades, so that markers lie extremely close to one another. For example, if x goes from 1 to 1000, there are 10 points in the first decade 90 in the second, and 900 in the third. You want to have, say, 3 points per decade instead.
I see two ways to solve the problem. The easier one is to use differently colored lines instead of different markers. Thus, you don't sacrifice any data points, and you can still distinguish everything.
The second solution is to create an unevenly spaced index. Here's how you can do that.
%# create some data
x = 1:1000;
y = 2.^x;
%# plot the graph and see the dots 'coalesce' very quickly
figure,loglog(x,y,'.')
%# for the example, I use a step size of 0.7, which is `log(1)`
xx = 0.7:0.7:log(x(end)); %# this is where I want the data to be plotted
%# find the indices where we want to plot by finding the closest `log(x)'-values
%# run unique to avoid multiples of the same index
indnew = unique(interp1(log(x),1:length(x),xx,'nearest'));
%# plot with fewer points
figure,loglog(x(indnew),y(indnew),'.')