< Back to Research

Autonomous [ESG] Portfolio Construction

Raphael Fiorentino
9th March 2020 - 7 min read

Butterwire’s latest feature enables users to automatically generate portfolios that:

  • meet their own investment strategy
  • that are built from their own stock universe
  • that match their target number of positions
  • have the highest forecast information ratio.

With this new feature, users will be able to rapidly construct tailored portfolios to both meet external (e.g. new client) and internal (e.g. reference portfolio) needs.

Below, we outline how we used the feature to build Global concentrated ESG portfolios of about 30 holdings.

The whole process takes but a few minutes and consists of straightforward steps:

  1. Short-list all the stocks that are eligible for portfolio inclusion (40 to 200 stocks)
  2. Select the target number of holdings and investment strategy (value, growth, etc.)
  3. Enter portfolio details (name, benchmark, currency, value) and press “Proceed”.

The portfolios are constructed by maximising the number of independent bets for a given number of holdings in a way that is consistent with the constraint set by the investment strategy. An overview of the principles underpinning the application is available at the end of the note together with an illustration of outputs for the various strategies.

Shortlisting stocks

Option 1: Directly via “Explorer”

The screenshot below reflects the criteria and filters that we decided to apply in order to narrow the 5,973 available stocks in the application down to below 200 (181 in the present case). This was done by going to the Explorer menu option. We excluded all stocks unlikely to be SRI-compatible, those with poor tradability, or with a red flag, or with an exit or check thesis alert, or with very low (fitness, value, momentum) fundamentals. Also excluded were all banks and insurance stocks, as well as stocks with specific macro profiles (e.g. stocks with a high sensitivity to economic cycles, to monetary policy, and to rising inflation expectations / positively correlated to the direction of commodity prices).


Next step just consists on pressing the “Extract” button at the bottom of the page.


Note: Instead of manual entries of filters and criteria, preset screens may also be programmed for instant shortlist updates (please contact us if interested).



Option 2: Via CSV Upload (for an even more refined shortlist)

The above selection can also be downloaded as a CSV file. This allows you to refine the list.


Then go to the Portfolio menu and select Extract Portfolio.

Portfolio Options

“Extract Portfolio” allows to upload a csv file containing your shortlist. Any csv file can be used as long as it contains one column with the label id or bwid (e.g. EU_0001 for Nestle).



Selecting number of holdings and investment strategy

The application then calculates a set of portfolios from the highest recommended number of positions (whose value will depend on the number and characteristics of the stocks shortlisted) to a minimum of 15.


The default investment strategy is “unconstrained”, which seeks the best overall mix of fundamentals regardless of style preference. This can be changed to any of the pre-set investment strategies by clicking on the “Portfolio Mandate” drop-down menu.


Click on the number of holdings closest to target to reveal the corresponding holdings and then press “Proceed” to confirm your choice.



Entering portfolio details

Upon pressing “Proceed”, a page appears that is the same as the one that follows the upload of an existing portfolio, allowing to specify the portfolio name, start date, benchmark, currency and value at inception.



Comparing outputs between strategies

The portfolios generated do vary significantly with the selected investment strategy. Below are the screenshots of 3 possible “green” portfolios (target of ca. 30 holdings) extracted from our shortlist of 181 stocks. The unconstrained portfolio has more provocative industry and country exposures (large underweights in US and Staples vs. overweights UK-Japan and Healthcare) relative to the value and recession-resilient portfolios (albeit the former is materially overweight consumer discretionary and the latter overweight industrials, specifically defense and services). There is less than 33% overlap between the unconstrained portfolio and the two others.


Testing each strategy enables to visualise the effect of majoring on a portfolio characteristic relative to an unconstrained mandate. As expected the unconstrained portfolio has the best overall mix of fundamentals, however it concentrates many momentum scores around its 7.3 / 10 average (lack of dispersion of either momentum or value scores may expose the portfolio to a synchronised sell-off). The same applies to the momentum strategy (as highlighted by the red borders around the 8.5 value in the table below). As for the recession-resilient strategy, in the present case it doesn’t deliver a materially higher recession-resilient score, entails a similarly high tracking error than the above two strategies, and has a lower expected alpha. The other five strategies yield similar tracking errors (5.5%-6%) and entail different compromises: the value portfolio scores lower on fitness/growth and momentum, the growth portfolio lower on value and volatility, the income and low volatility ones lower on alpha and recession-resilience, while the quality portfolio shows middle of the road characteristics.


It is important to note that the above results are highly dependent on the initial shortlist (specifically the dispersions of characteristics therein) and, crucially, on trailing market conditions: stocks with high momentum scores will tend to exhibit high recession-resilience scores if markets have been pricing in a recession over the past few weeks, which in turn will drive some convergence of portfolio characteristics between the unconstrained, recession-resilient and momentum strategies.


Addendum: Portfolio Construction Principle

The autonomous portfolio construction feature derives from a Principal Component Analysis (PCA) of the selected stock universe’s daily total returns (price changes + dividends) over the past year, with nearer days being made to contribute more information to the overall returns variance analysed (e.g. the nearest 90 days contribute 1.35x more information than the previous 90 days and so on). It exploits the properties of PC Axis whereby each of the, say 181 stocks, represents a long or a short position of varying weight in each of the 181 composite portfolios generated by the PCA, and in such a way that the % of total variance explained goes from highest (Axis 1) to lowest (Axis 181):

  • The first PC axis represents a large % of the stock universe’s total variance (31% in our case) as it captures overall market variance (i.e. systemic risk) – stock weights on this axis are typically clustered (around a positive but not statistically significant value) and as such it is not an axis that has much impact on the construction (exceptions may include very high beta and very low beta stocks, e.g. gold equities, in which case the engine will favour the latter)
  • The significant axes (eigenvalues > 1), 32 of them on our case, represent another 46% of the total variance. They each capture significant independent bets that reflect specific combinations of risk factors (e.g. one such composite portfolio may look very long Eurozone Healthcare stocks and short US Industrial stocks). We apply factor rotation (Varimax) to more easily surface the stocks primarily responsible for these significant contributions to variance. The engine then looks for ways to both preserve a high number of independent bets and minimise the ultimate portfolio variance. It does so by mitigating the number of stocks with significant (respectively positive and negative) loadings on each of these axes.
  • The insignificant axes (eigenvalues ~ 0), 118 in our case, explain in aggregate less than 10% of total variance and therefore mostly capture random noise. As such they may also reveal extremely co-variant, and therefore interesting stocks, to the extent that their weights are (positively or negatively) significant. For instance, the past share price evolution of two stocks with high weights of opposite signs on such an axis will tend to look very similar (e.g. Rio Tinto and BHP). Holding both would likely concentrate exposure and reduce the number of independent bets per holding. Conversely, when two stocks “load” heavily on such an axis and have the same sign, their past share price evolution will tend to look more (or less) like mirror images (e.g. Amundi and Steris). Holding both is likely to improve the risk profile of the portfolio.
  • The remaining 30 axis are disregarded