Using Economic Indicators with Eikon Data API - A Machine Learning Example

From this if we look at the Economic Indicators section we can see that over 4000 items are returned. So lets click on that to discover more.

We are now at the main filter page so we can filter from the 10 columns that are there. The most useful will probably be source, periodicity and perhaps most importantly Indicator Type. Please see the types listed below:

Multiple Economic Indicator Data Sources

Once we narrow down what we are targetting we can then go to the page view and then download a list of appropriate RICS to excel. From there we can copy and paste these as an instrument list for use in our API Call.

Real-time Economic Indicators

As mentioned earlier Eikon carries about 1800 real-time economic indicators. Lets see what we are taking about by looking at the economic monitor (type EM into Eikon Search) for say the US (filter on US). Notice here we can see the recent and upcoming economic indicators for the next week or so with a date as well as time, importance of release, polls to see what expectations are, as well as details about these polls such as min, max, median.

For realtime indicators lets make a simple call to the Data API get_data service. Here we are using the no longer maintained economic chain RIC '0#ECONALLUS'. These still work for backward compatibility but are no longer maintained from a content perspective. This can give us a programmatic version of the Economic Monitor App above. We are using the realtime quote model for economic indicators here and you can see this by typing a realtime RIC - say 'USADP=ECI' (US ADP payrolls) and hitting quote. If you hover the mouse over a particular number you will get the field name.

Real-time Indicator retrieval from the timeseries service is different from other RIC instruments in that we DO NOT request field types (eg Open High Low Close) and we set the interval to monthly. Lets see how it works using the ADP payroll number for the last 10 years of monthly releases. Note that I have also added 3 other RICs that seem to be derived from the same underlying RIC. These are the poll values associated with the monthly release (hence the lowercase p preceding the RIC) and then we have Median, Low & High Values (the =M,=L,=H respectively succeding the RIC)

123 rows × 4 columns

Non Real-time Economic Indicators

As mentioned earlier Eikon carries just shy of 400,000 non-realtime economic indicators covering a whole host of indicator types listed above. Once you have found the list that you wish to download - just copy and paste those into a historical timeseries call this time for Monthly German New Car Production and registration:

123 rows × 3 columns

Putting it all together - a basic machine learning sample

So given our new found economic indicator discovery workflow - lets try to put this data to work to see if we can come up with something of interest. In this case lets try to download all available auto industry indicators for all countries in the world and see if this can help explain movement in some global automakers. In order to do this we will be using an XGBoost model which a gradient boosted decision-tree class of algorithm.

From the Economic Search app I have identified a list of global economic indicators from the automobile industry and have downloaded these rics to an excel file - which I can now read into pandas and also use in my get_timeseries call.

['IDMBKY=ECI',
'IDCARY=ECI',
'USCARS=ECI',
'ESCARM=ECI',
'ESCARY=ECI',
'ITCARM=ECI',

...

'aJPPNISS',
'aCATRKSBSP',
'aUSAUIMMXP',
'aGBACECARP',
'aLUNCAR',
'aISNVEHREG/A',
'aROACECARP']

Importantly, we now want to create some lags for our data to see if there is any lagged effect for our data.

Adding Target Variables or Labels

So we now have our feature matrix with all our raw and lagged data. Now lets add some target variables or labels. We will be adding two types of label that I think encompass many of the use cases - firstly a timeseries of monthly closing prices as a label - and secondly a company fundamental monthly timeseries - in our case the IBES consensus revenue estimate for FY1. You will see its very straightforward to download and then concatenate these types of label data to your feature matrix. Here we use BMW as an example as it is a global automaker.

121 rows × 2 columns

Next we need to standardise our data before we feed it to the XGBoost routine and that means both the factor matrix and the labels. In our case we will use a simple to calculate z-score. As you will see this is applying to all our current feature and label concatenated frame.

Next we need to make sure we identify the right features to pass to the XGBoost model - we only wish to pass the zscore features.

Index(['IDMBKY=ECI_zscore', 'IDCARY=ECI_zscore', 'USCARS=ECI_zscore',
'ESCARM=ECI_zscore', 'ESCARY=ECI_zscore', 'ITCARM=ECI_zscore',
'DECARM=ECI_zscore', 'DECARY=ECI_zscore', 'GBCARY=ECI_zscore',
'ZAVEHY=ECI_zscore',
...
'aISNVEHREG/A_lag_3_zscore', 'aISNVEHREG/A_lag_4_zscore',
'aISNVEHREG/A_lag_5_zscore', 'aISNVEHREG/A_lag_6_zscore',
'aROACECARP_lag_1_zscore', 'aROACECARP_lag_2_zscore',
'aROACECARP_lag_3_zscore', 'aROACECARP_lag_4_zscore',
'aROACECARP_lag_5_zscore', 'aROACECARP_lag_6_zscore'],
dtype='object', length=1813)

We are now getting into the machine learning section. So we need to split our data in train and test sets.

We also need to a function to create our feature and label sets and then we go ahead and create our X_train (training features set), y_train (training label set), X_test (test features set), y_test (test label set).

Finally I usually just do a date check - to see all dates are properly aligned.

2010-03-31 00:00:00 2010-03-31 00:00:00 2014-12-31 00:00:00 2014-12-31 00:00:00

2015-01-31 00:00:00 2015-01-31 00:00:00 2020-03-31 00:00:00 2020-03-31 00:00:00

Define and fit XGBoost model

Finally we start defining oour XGBoost model (in our case we use a simple vanilla setup) - then we pass our training set and test set to fit our model. Once we have done that we can have visualise our feature importance as a plot.

XGBRegressor(base_score=0.5, booster='gbtree', colsample_bylevel=1,
colsample_bynode=1, colsample_bytree=1, gamma=0,
importance_type='gain', learning_rate=0.1, max_delta_step=0,
max_depth=3, min_child_weight=1, missing=None, n_estimators=1000,
n_jobs=1, nthread=None, objective='reg:squarederror',
random_state=0, reg_alpha=0, reg_lambda=1, scale_pos_weight=1,
seed=None, silent=None, subsample=1, verbosity=1)

Finally - we can use our model to create an ML_Prediction for our label set given our X_test set and then compare that to actual observations of the label set.

Conclusion

In this article we have discussed the economic content available in Eikon. We talked about real-time and non-realtime economic indicators and how you can discover and use this data in your workflows using our search discovery tools. We also showed an example of real-time indicator monitoring - similar to the Economic Monitor (EM) app in Eikon.

Then we moved on to an example which puts a lot of this into practice using a simple machine learning example. We found a group of automobile industry data using our search discovery tools and then saved those down as a spreadsheet. The we used those RICs to download the timeseries for these economic indicators. We created some new (derived) features by creating 6 months worth of lagged data. Once we had completed our raw feature matrix - we showed how to get two types of typical label data - a price series (in our case monthly BMW closing prices) and also an example of some fundamental data for BMW (in our case Monthly IBES consenus Revenue estimates). We then standardised our data using z-scores, a very common method followed by splitting our data into training and test sets and created our feature and label sets for both. We then defined a simple XGBoost model and handed our data to it to fit. Once that was complete we then created a model prediction for our label set (in our case BMW monthly closing prices) and then visualised this is a chart.

The model was able to generate some kind of prediction - which doesn't seem that compelling - but is not entirely without merit. The model is broadly predicting lower prices for the stock based on our economic indicator factors and this has broadly been evidenced in the test set.

Whilst a simple example, I hope this article has demystified this type of workflow and provided you with a practical codebase to explore further.

Further Resources for Eikon Data API

• Overview
• Quick Start
• Documentation
• Downloads
• Tutorials
• Q&A Forums

For Content Navigation in Eikon - please use the Data Item Browser Application: Type 'DIB' into Eikon Search Bar.

Download Notebook from Github