You can do that in Research, and I heard they were planning on adding historical fundamentals to the backtester/live trader, but I don't think it's happened yet.

Hello,

Here is an example algorithm that compares current earnings per share to last earnings per share. With the current implementation, the solution is a bit of a hack, but as we improve pipeline, it will become easier to do this. For now, this implementation should get you what you want -- especially the EPSChange custom factor.

Please let me know if you have any questions.

Oh duhh, of course, could use pipeline for this :)

Hello,
Thanks for your reply.
@ Jamie I tried using pipeline as you suggested but the backtest doesn't work, maybe because of long computation time in factor EPSchange.
Some one has an idea to deal with this issue ?

Thanks
idriss

It looks like you were trying to increment context.days before having initialized it. As well, in regular_allocation, you were looping through sids in context.output.index instead of data. I've attached a working version of your code with the fix. I hope this helps!

Hi Jamie,

unfortunately it's not possible to use np.unique(.) to retrieve the last 4 quarters, because it could be that different quarters have the same values.

I've implemented the following logic. It's computationally correct if tested standalone, but within the backtester a TimeoutException always occurs... may the loop is too slow:

days_in_quarter = 70  
NA = object()

def retrieve_ttm(column):  
    ttm = []  
    i = 0  
    j = 0  
    ttm.append(column[i])  
    for previous, current in zip(column, column[1:]):  
        i += 1  
        if np.isnan(previous):  
            previous = NA  
        if np.isnan(current):  
            current = NA  
        if previous != current:  
            if i - j > days_in_quarter:  
                ttm.append(column[j])  
            ttm.append(column[i])  
            j = i

    if len(ttm) < 4:  
        ttm.append(column[i])

    return ttm[-4:]


class RevenueTTM(CustomFactor):  
    inputs = [morningstar.income_statement.total_revenue]  
    window_length = 252

    def compute(self, today, assets, out, revenue):  
        out_list = []  
        for i in range(revenue.shape[1]):  
            column = revenue[:,i]  
            ttm = retrieve_ttm(column)  
            out_list.append(ttm)  
        all_revenues = np.transpose(np.array(out_list))  
        revenue_ttm = np.sum(all_revenues, axis=0)  
        out[:] = revenue_ttm[0]  

The full algorithm is attached.
How may I implement this feature in a more efficient was? Thanks!

I wonder if one could see the history of xxx_as_of and pick out the values when that changes...

Well, I'm trying to do that... the code above iterate over the last 252 day (a trading year) and store the value on each change or when there isn't any change for more than 70 days (meaning the value in the current quarter is the same as in the previous one).
The problem is than the loop takes too long for Quantopian and a timeout occurs.

I cannot yet figure how to overcome this issue and hope in the help of the community.

I attach now the full algo again because I fear the source code wasn't available in my previous post:

from quantopian.algorithm import attach_pipeline, pipeline_output  
from quantopian.pipeline import Pipeline  
from quantopian.pipeline import CustomFactor  
from quantopian.pipeline.data import morningstar  
import numpy as np

days_in_quarter = 70  
NA = object()

def retrieve_ttm(column):  
    ttm = []  
    i = 0  
    j = 0  
    ttm.append(column[i])  
    for previous, current in zip(column, column[1:]):  
        i += 1  
        if np.isnan(previous):  
            previous = NA  
        if np.isnan(current):  
            current = NA  
        if previous != current:  
            if i - j > days_in_quarter:  
                ttm.append(column[j])  
            ttm.append(column[i])  
            j = i

    if len(ttm) < 4:  
        ttm.append(column[i])

    return ttm[-4:]


class RevenueTTM(CustomFactor):  
    inputs = [morningstar.income_statement.total_revenue]  
    window_length = 252

    def compute(self, today, assets, out, revenue):  
        out_list = []  
        for i in range(revenue.shape[1]):  
            column = revenue[:,i]  
            ttm = retrieve_ttm(column)  
            out_list.append(ttm)  
        all_revenues = np.transpose(np.array(out_list))  
        revenue_ttm = np.sum(all_revenues, axis=0)  
        out[:] = revenue_ttm[0]


def initialize(context):

    # Create the pipe  
    pipe = Pipeline()  
    attach_pipeline(pipe, 'Revenue-TTM-Example')

    revenue_ttm = RevenueTTM()

    pipe.add(revenue_ttm, 'revenue_ttm')  
    # Create and apply a filter representing  
    revenue_ttm_top_100 = revenue_ttm.top(100)  
    pipe.set_screen(revenue_ttm_top_100)

def before_trading_start(context, data):  
    context.output = pipeline_output('Revenue-TTM-Example').sort(['revenue_ttm'], ascending=False)

    update_universe(context.output.index)

def handle_data(context, data):

    print "SECURITY LIST"  
    log.info("\n" + str(context.output))  

Yeah you definitely shouldn't do any looping in a pipeline factor. You must do it vectorized.

I re-implemented without loops but the Timeout still occurs :-(

Here the new implementation:

days_in_quarter = 70  
num_of_quarters = 4

def retrieve_ttm(array):  
    # Differences between the next and the current value, after havind convertet NaN and Inf to numbers  
    diff = np.diff(np.nan_to_num(array))

    # Track the indexes where the difference isn't zero: a change has occurred. Shift the indexes by one.  
    idx = np.where(diff != 0)[0] + 1

    # Add zero as the first index (by definition)  
    idx = np.insert(idx, 0, 0)

    # Look for periods without changes longer than 'days_in_quarter'  
    dup = np.diff(idx) > days_in_quarter

    # Add an index in order to duplicate the previous value  
    idx = np.append(idx, idx[dup] + days_in_quarter)

    # Sort the index for consistence  
    idx = np.sort(idx)  
    return array[idx[-num_of_quarters:]]  

Hmm that's too bad :(