- Text Mining of Corporate Responsibility Reports
- Operational Risk And Controls Assessment (RCSA) Visualisations
- Large Project Gantt Style Visualisations
- Visualisations based on the Gapminder dataset
I gave a talk to Microsoft Student Partners about incubating machine learning into a large organisation, in particular into a big bank. The slide deck is here. I mentioned several events and articles during the talk. Here are the links.
Rohan Kopparapu (4th year UCL) and Microsoft Student Partner wrote a blog post about the UCL Data Science Student Challenges.
The video recordings of the two recent talks about building a data-driven culture in an organisation are Transforming a Museum to be data-driven using R – Alice Daish and
From nothing to something, a local government journey- John Kelly.
The podcast of the BBC Radio 4 interview with Anthony Jenkins, ex-CEO of Barclays, and others about how fintech and ML are disrupting banking is available here.
I suggested some London-based meetups
https://www.meetup.com/London-Business-Analytics-Group/ (LBAG, my one)
The Economist article on the change of culture at Microsoft and its focus on cloud and AI is here.
The abstract for the talk is below.
Banking On Machine Learning
Many of the audience will have recently attended a hackathon at UCL. Over a weekend participants applied Azure Machine Learning to financial datasets supplied by Bloomberg and derived some insightful results – a fantastic achievement. This was for many their first taste of a hackathon and data science. However, for many organisations, getting to the stage where they are doing hard data science as part of a viable project is a long way into the journey of introducing and successfully taking advantage of ML.
This talk looks at the experience of starting ML at a few organisations then focuses on the challenges faced by big diversified banks. We look at why banks are rushing to embrace ML and consider both the opportunities and threats that ML poses. We’ll consider the steps to introduce and incubate ML within the bank; through awareness, education, crowdsourcing, mobilisation (including hackathons) and finally implementation of ML as the new normal.
We’ll conclude that the successful introduction of ML into an organisation is about communication, culture, change, marketing, business analysis… and finally the hard data science of the sort that we saw in the hackathon.
I gave the March round-up talk at the London Power BI User Group last night, before John Kelly’s inspiring talk, without the benefit of slide deck or demos due to the AV issues. So here is my slide deck and links of the events and resources I mentioned during my talk. And I have posted 8 minute video to demo the Power BI desktop February and March updates on youtube, and also on OneDrive.
The details of the SQLBI ‘Mastering DAX Workshop’ training in London on March 22-23 are here.
The PASS Business Analytics Marathon on March 29th features two Power BI talks from London PUG organisers; Prathy Kamasani at 6pm and David Moss at 7pm. Register here.
The PBI motion scatter diagram of the Gartner Group Magic Quadrant of the last several years is here.
Now that we can theme our colours, the color brewer site is a very useful resource for selecting good palettes of categorical, sequential or diverging colours that forms best practice.
The Power BI team video of the March desktop update is here.
The data storytelling course by Albert Cairo is here.
The Economist produce wonderful charts on a daily basis in their Graphic Detail section of their website.
This page contains the links and references for my talk “Using R To Clean And Transform Small Data” at the PASS BA Marathon on 14th December 2016.
The video recording and slide deck are here.
The Microsoft Professional Program Certificate in Data Science has two R courses. These are the Introduction To R For Data Science and Programming With R For Data Science. I also recommend the DataCamp online R courses.
The London Business Analytics Group (LBAG) meetup page where you can see upcoming events is here and the videos of some of the past events are here. The London Power BI User Group (PUG) meetup page is here. Mark Butler’s blog of recent PUG and LBAG events is here.
The third way of using R in Power BI is to create an R Script visual. This is especially useful when we want to create a visual such as small multiples that is not available (yet) in the suite of standard or custom PBI visuals.
Image that we have a daily time-series for December 2015 for the profit and loss (P&L), and we want to show a trellis of bar charts; one chart for each region (Asia, EMEA and US) and one for a Volcker flag (which is either In or Out). So we want to show six charts in all. We also want to show a bar on each chart in a different colour for each of three clusters. Given a data frame named dataset with date, Region, Volcker, Cluster and P&L columns, we can create a this in R using the ggplot2 library with the following code.
Once we have written and tested this code in RStudio, we can create a R Script Visual. The RScript visual provides a dataframe (R’s equivalent of a table) named dataset. This contains a column for each of the fields in the Values well – in this case Region, Cluster, VolckerFlag, Date and P&L.
We can copy our R snippet into the bottom of the R Script Visual code window.
Once done we can see our chart in Power BI.
The second way of using R in Power BI is as a step in a query in the Query Editor. Imagine we have a Task query that has a TaskOwner column that contains a comma separated list of one or more owners – like the snapshot below.
Our objective is to create a new query that contains the mapping of tasks to owners – this requires us to split the TaskOwner column and create a row for each task and owner.
Since we are creating an additional dataset, let’s start by duplicating the Task query and rename to TaskOwnerMap. We can now insert a R step (rightmost icon on the Transform ribbon below).
This opens a dialog where we can paste our R code snippet.
The code above uses the functions from the tidyr package. The separate function splits the TaskOwner columns into several columns, one for each owner. The gather function then unpivots these columns to create an attribute and value column. The select chooses the columns required for our final dataset. (It’s best to develop, debug and test in the RStudio or other development environment beforehand.)
This transforms the data as required – we’ve named our result dataframe task.owner.map and we expand this in the next step to see the columns.
This particular example (splitting a column) is more easily accomplished directly in the Query Editor than using R but I wanted to show a simple case to explain the mechanics of the process.
There are three ways to use R in Power BI
- Load a R Script as a data source
- Add an R Script as a QE step – introduced in the July 2016 update
- Plot a R Visual
In this article we’ll look at the first of these. The subsequent articles will look at the other two ways. We’ll use as an example in a typical reporting and visualisation challenge taken from the area of report progress of task to meet certain regulatory requirements – in this instance the 11 principles for good aggregation and reporting of data issued in by the BCBS – details here if you are interested.
Imagine the source data looks like this.
This shows a row for each task with some properties (Key, Name, Start Date, End Date). It also shows which task are applicable for which regions (denoted by the x in the cell). The set of regions is implicit in the column names. It also shows the owners of each task – if there is more than one owner, these are in a comma separated-list. In a similar way, it also shows the principles covered by each task. There are 11 principles and we know this list will only have integers between 1 and 11. We also have on a separate sheet the details of the principles
This is not a good structure for reporting and analysis. We need to tease out the data into separate, related tables –
- TaskPrincipleMap (to hold the many to many relationships between Tasks and Principles),
- TaskRegionMap (again to hold the many to many relationships between Tasks and Regions)
and combine this with the Principle table to arrive at a data model that looks something like this
The R script below uses the tidyr and dplyr packages to perform a sequence of data transformations to achieve exactly this. These include projecting columns (select), filtering rows (filter), splitting a comma-separated column into individual values (spread) and unpivot (gather). The result is a set of data frames (R’s equivalent of a table) that correspond to the tables required.
We can paste this RScript into the “Load R Script“ window.
One of the nice features of this is that you can then choose exactly which dataframes created by the script to import – and that you can load several dataframes not just one.
This then loads our data into the Query Editor below – each selected R data frame becomes a separate query. (in the snapshot below, I have renamed the queries from R naming convention to more typical table / query names e.g. from df.task.principle.map to TaskPrincipleMap)
In this case, we can do these operations directly in Power BI of course – but R has many advance statistical capabilities that are not in Power BI; decision tress, clustering, correlations. Some of these are shown in the R Script Showcase here.
Operational risks results differ from other types of financial risk results such as market or credit risk. The latter are defined in numerical ratio terms – examples include profit and loss, value-at-risk and exposure. We can add these measure together naturally.
However, operational risk results are usually expressed in ordinal terms. For example, an operational risk and controls self-assessment (RCSA) estimates the inherent risks, assesses the controls that are in place to avoid and mitigate these risks and then estimates the residual risks after these controls are applied. The effectiveness of each control is also estimated – typically on a scale from adequate, inadequate, vulnerable and worst-of-all not controlled.
There are a set of 7 industry standard well defined event type categories; internal fraud, external fraud, employment practices, clients & products, damage to physical assets, business disruption and execution & delivery. Each of these is broken down into sub-categories. For example, internal fraud splits into unauthorised trading, bribery and others. Furthermore, many parts of the bank submit a RCSA assessment and these may be organised into a hierarchy.
Risk Managers want to view these as heat maps. They would view initially at an aggregated level, for example, at event type category and across all regions. The aggregated values should be the “worst case” of the detail values so if a department has estimated low for misappropriation of assets, medium for mismarking and high for bribery then the internal fraud aggregated score would be high.
The snapshot below is a typical heatmap of RCSA scores, mocked up in Excel. The columns list the inherent risks, mitigating controls these residual risks. These are broken down by region. The rows list the standard operational risk event types
The data for this heatmap is a table – first few rows below.
Let’s build this heatmap in Power BI Desktop. Our first problem is that these are ordinal results. L, M, H, VH values represent low, medium, high and very high scores for risks. Similarly, A, I, V, NC, values represent adequate, inadequate, vulnerable and not controlled for effectiveness of controls. So we add another dataset that helps us map these ordinal scores to a numeric weight to indicate the order of these scores. The actual values of those weights don’t matter, just the relative values.
We’ll merge these two datasets in the Query Editor to get a combined data set like this below
We then create two measures
WorstCaseWeight = MAX(RCSAData[Weight])
This measure aggregates as worst-case manner as required.
WorstCaseScore = SWITCH([WorstCaseWeight], 10, “L”, 20, “M”, 30, “H”, 40, “VH”, 110, “A”,120, “I”, 130, “V”, 140, “NC”)
This measure maps our worst case weight into the appropriate ordinal score. Power BI does not allow a string field in the values well so we can’t add our score values directly but it does allow any measure – even if that measure returns a string. This is very useful for us.
Now we can visualise our heatmap both at an aggregated and a detail level.
This gives us basically what we need with the exception of conditional formatting of the cells. We have a heat map without the “heat” which is unsatisfactory. What we need is to be able to map these categorical scores to colours and apply these colours to the cell background. This is not yet possible for the matrix visual, or indeed for any Power BI visual but I hope that this will be a features of upcoming update – or that someone will build a heat map custom visual.
The Power BI Desktop May Update introduced conditional formatting of numerical columns in tables. The snapshot below shows a table with two columns, one of which has had conditional formatting applied.
The conditional formatting dialog is a context menu on numeric columns in the values well
You can set the colours at the end points of the colour scale, the minimum or maximum values from a colour picker but also choose a customer RGB colour value (for example, #FF0000 representing pure red). There is a lot of choice but very little guidance on the best combination of colours. Fortunately the colour brewer website can help here by suggesting a few good colour schemes. This site was developed by Cynthia A. Brewer in the Geography department of Pennsylvania State University.
The simplest approach is to take the RGB values of the first and last colours in a scheme and enter these as the custom colours for the minimum and maximum values.
For example, the PL column in the table above used # edf8fb as the colour for the minimum value and # 006d2c as the colour for the maximum value
The conditional formatting dialog allows to choose a diverging colour scheme. This is useful for a range of values centred around zero where you want to emphasize large absolute values either positive or negative. For example, on a column of profit and loss values it is useful to highlight those periods with either large profits or losses.
This is the first appearance of conditional formatting and there are of course some improvements that suggest themselves. It would be helpful to be able to
- To be able to set some defaults for personal, team or organisational colour schemes – currently every column has to be set up manually
- To have a set of discrete colours mapped to ranges (as is provided by the colour brewer site)
In the longer run it would be very useful to have conditional formatting based on a categorical measure as well as a numeric range. For example, imagine a DAX calculation that returns either “Breach” or “OK” – we would like to highlight the breaches. Of course we could fake this now by our DAX calculation return -1 or 1 for OK and Breach respectively.
The Power BI Desktop May Update introduced Quick Calcs. In fact, there is currently only one calculation currently, percentage of grand total. However, it shows the direction of travel – I assume that there will be other calculations added in future months; e.g. percentage of parent and time related calculations; moving average, year-to-date and so on. It removes a stumbling block for many (perhaps most) users so that they can access typical common calculations; if you know DAX then you can write a concise DAX function for any of these calculations, but many users will not want to learn DAX. Quick Calcs are useful with the new tooltips feature, also introduced in the May update
You can create a quick calc from a context menu on a measure placed in the Values or new Tooltips area.
The ‘Show value As’ drop down in the Quick Calc dialog provides only one option ‘Percent Of Grand Total’
The snapshot below shows the tooltips area with two measures; the ‘% GT PL’ generated by the quick tooltip and a manually created DAX measure for comparison ‘% GT PL DAX’ that generates the same result. This has the formula
% GT PL DAX = [PL]/ CALCULATE( [PL], ALL(FirmHierarchy) )
And the tooltip shows the PL measure (as usual, since in the values area) but now also the two measures in the tooltip area.
This is the first appearance of Quick Calcs and there are of course some improvements that suggest themselves. It would be helpful to be able to –
- format the result, just like you can a normal DAX measure
- rename the result, again just like you can a normal DAX measure. The name is set to ‘% GT <original measure>’ and is sensible but of course may not fit in with the pattern that you use to name measures
- put the calculation into the fields list so it can reuse it in other visuals
- see the DAX underlying the quick calc; this would be a good way of learning some practical DAX