Using Stanford Corenlp, a natural language software, designed to capture the sentiment of sentences expressed in a natural language (e.g. English), we have developed procedures for analyzing investors’ emotions and beliefs concerning the short-term evolution of economic and market conditions.
These views may be found in public and popular sources, such as online news and social networks (e.g. twitter) or online news agencies (e.g. Bloomberg). The results from such a “deep learning” application are then introduced as additional inputs in the part of our investment process, in which we estimate shifts in market expectations about the realization of risk factors.
This algorithm is also used to analyze market’s sentiment about individual stocks. The outcome of this analysis is then mapped into a numerical scale, ranging from -100 (extremely negative sentiment – sell disposition) to +100 (exceptionally strong sentiment – buy disposition). Based on these rankings we construct a list of individual American and European stocks, ranging from the hottest to the most unpopular ones. This list is updated regularly and is available to our clients upon request.
A separate but complementary approach for stock selection is based on a set of financial metrics that serve as criteria for picking the stocks with the highest probability of over-performing in the medium term. This approach comprises of the following three steps:
- From the initial universe of the most widely used financial metrics (e.g. P/E, P/BV, ROE, Earnings growth, Earnings Forecasts, Analysts Recommendations etc.), we identify the subset of those metrics which contain the most relevant information for the future behavior of stock price. This subset is determined using “big data” methods which allow for the uncovering of hidden (and time varying) correlations in large and varied data sets.
- We determine the weights by which each metric enters the selection algorithm. These weights are not constant over time but changing as a function of evolving economic conditions. The exact form of this function is determined by methods of “deep learning” which are designed to extract an underlying structure from a set of raw data.
- Having defined the selection criterion (in steps 1 and 2) we use it to rank the stocks of SP500, Eurostoxx 600 and Swiss SMI. The end result of this procedure is a list of the most preferred stocks which is available to our clients upon request.
In the context of our asset allocation decisions, anticipating the moves in major exchange rates, especially the dollar-euro rate, plays a significant role. To this end, we have developed the following procedures for exchange rate forecasting:
- Forecasts through Structural Econometric Modelling: Specific forecasts for the dollar-euro rate are obtained by means of Assetwise’s global econometric model. In this framework, the dollar-euro rate is modelled in terms of growth, interest-rate and inflation differentials between the US and Europe. The coefficients of this equation are not assumed to be constant over time, but are evolving over time in line with the changing economic conditions.
- Survey of Analysts’ Forecasts: Every month, the exchange rate forecasts of major analysts are surveyed and analyzed using Assetwise’s decision support system. The average forecast is then computed and compared with the one produced by Assetwise’s econometric model. Special emphasis is paid on examining the diversity of analysts’ forecasts. An index measuring the degree of agreement/disagreement of these forecasts is constructed and used as an additional input in our overall exchange rate assessment.
- Trend Detection Algorithm: Apart from fundamental causes, exchange rates are driven by the so-called technical factors, including leverage, carry trade and such. These additional forces tend to create local trends in the temporal behavior of exchange rates. Methods for detecting such trends, soon after their creation, may offer valuable insights for the short-term evolution of exchange rates. To this end, we have developed such a trend detection algorithm which is part of AssetWise’s integrated decision support system.
Hedge funds express conviction in companies they hold by sizing them up in their portfolios. If a stock accounts for a significant portion of a manager’s portfolio, it is considered an investment with conviction. To disentangle between conviction and consensus (i.e. companies that belong to the portfolio simply because they constitute large components in various benchmarks), we adjust portfolio weights to reflect true conviction.
The first step is to identify Hedge Fund managers with superior skills. The initial universe is constrained to include funds with a long-only strategy and long lock-up periods, in order to keep those portfolios that consist of high conviction picks with a long term focus. We then construct a portfolio of stocks with the highest number of managers invested with conviction. Stocks in this portfolio represent the cumulative stock selection and position sizing skill of the hedge fund industry.
The bond platform generates bond portfolios that comply with the constraints of institutional investors while minimizing the default probability. In the first step of the process, the platform screens the bond universe and collects bonds that satisfy user-specific criteria.
The second and most important step consists of building the optimum portfolio subject to various constraints, at the portfolio level. The multiplicity of the platform's features allow the user to build portfolios that satisfy his specific goals.