by Bernardo Bravo, Senior Data Scientist, Yewno,
and Ali Limon, Senior Data Scientist, Yewno


In this white paper we present how investors can leverage the Yewno’s Knowledge Graph Concept Exposure data to create investment strategies based on an individual concept used as a standalone signal, in other words, a concept-based factor strategy. 

The strategies select securities from the SPY1 ETF constituents, based on differing degrees of exposure to the concept ‘Partnership’ extracted from news content. We present a long short weekly-rebalanced strategy that consistently outperforms the benchmark on a risk adjusted basis during the time period considered, by as much as 7% per annum on net return.

Introduction and rationale

The use of Knowledge Graph approaches is growing for applications in finance.  This data framework provides a natural representation of the dynamic relationship among companies and news events. In particular, Partnership announcements in the news tend to be positive for the performance of companies involved and even companies profiting indirectly from it.

In this use case we leverage the Yewno Concept Exposure news scores to construct a factor representing the degree of exposure of companies to the concept ‘Partnership’. Specifically, we use the Aggregate Concept Exposure score which is a linear combination of the other Concept Exposure scores: contribution, pureplay, centrality and similarity2

The Aggregate score places a higher weight on contribution and pureplay which are derived from direct mentions between the concept ‘Partnership’ and the company concepts, whereas the centrality and similarity scores have a lower weight but include information of the entire graph incorporating indirect connections. 

Data Overview

Concept Exposures API query

The data used in this exercise was queried from the Yewno Concept Exposures API3, for news sources from Jan 2017 to April 2021 for companies in the SPY ETF. The lookback window used as a parameter is 15d. The request below is an example of the API call required to download news exposure data for a specific date, with the destination concept being partnership, and only four companies in the SPY constituents are shown in the source concepts side for simplicity. The lookback window is 15d, meaning that the scores considered news articles published within the previous 15 days of the corresponding reference date.


“sourceType”: “news”,
“window”: “15d”,
“dateStart”: “2018-07-18”,
“dateEnd”: “2018-07-18”,
“source”: {

“concepts”: [



“destination”: {

“concepts”: [“de4763607526b857bb84ab7ccd2d5b36”]


“expandIsins”: false,
“expandParents”: false,
“filters”: {

“entitiesOnly”: false,
“conceptsOnly”: false


“fields”: [




‘Partnership’ Strategies

Long only strategies

The concept exposure scores were queried with a reference date corresponding to every Wednesday to place hypothetical trades on the following Thursday. For each week we sort the companies by their Aggregate Exposure score to the concept ‘Partnership’, guaranteeing that at least 30 securities are included in each portfolio.

We create ten portfolios by decile weighted by the corresponding Aggregate score. The cumulative returns are shown in Figure 1. The decile of highest exposure A_D1 includes the securities with the relatively higher Aggregate score to the concept ‘Partnership’. The strategy that outperforms the benchmark is A_D10, the bottom decile with securities with a relatively lower score.

The Aggregated news exposure score reflects the extent to which companies are either directly or indirectly connected to the concept ‘Partnership’ in the Knowledge Graph induced by news mentions. In particular, the bottom decile of the Aggregate score usually includes companies with indirect connections to the concept as opposed to direct mentions, which would be more characteristic of the top deciles.

Figure 1: Cumulative returns on long only strategies for concept ‘Partnership’

Sources: Factset, Morningstar, Yewno

Figure 1: Cumulative returns on long only strategies for concept “Partnership”


In Figure 2 we present the performance metrics of the bottom decile strategy, over the whole period the strategy outperforms the benchmark by 3% annualized return. Moreover, despite the higher volatility the strategy outperforms the benchmark by risk/return ratio of 0.83 compared to 0.77 for the benchmark. The average daily turnover of the strategy is 40%.

We also show the top average holdings in the last year, observing that some companies such as Vornado Realty Trust, Welltower Inc, DaVita Inc were mentioned in the news in the previous year explicitly forming part of some Partnership. However, not all companies in the holdings had explicit news mentions with the concept, suggesting an indirect connection.

Sources: Factset, Morningstar, Yewno

Figure 2: Returns of bottom decile strategy on concept ‘Partnership’ and top holdings



Long / Short Strategy

Based on the decile strategies described above we constructed a long/short strategy in which we initiate a long strategy using the bottom decile (A_10) by 160% and initiate a short strategy with the top decile (A_1) by 60% (L/S 160/60). 

The average daily turnover of that strategy is 80%, also the volatility was higher and the max drawdown is worse than the benchmark but the strategy still remains superior with a Sharpe ratio of 0.96 compared to 0.77 for the benchmark.

Sources: Factset, Morningstar, Yewno

Figure 4: Performance metrics of the L/S strategy on ‘Partnership’ concept.


In this example we constructed a profitable trading strategy using Knowledge Graph Exposure Scores on news, despite the fact that the universe of securities in the S&P 500 is highly liquid and actively traded. Providing an indication that the Yewno Knowledge Graph alternative data exhibits signals that are not incorporated by the market.

The long short strategy constructed has a high volatility and large turnover but remains superior to the benchmark on a risk adjusted basis. On the other hand, the short of the top decile portfolio and the long of the bottom decile means that we are shorting companies with a direct connection with the “Partnership” concept, and going long companies that usually would have only indirect connections. This suggests that the data is capturing companies that would indirectly profit from the “Partnership” event and the market has not incorporated such information in the prices, further investigation on this is left for future work. 

1The SPDR S&P 500 Trust ETF (SPY) tracks the S&P 500 Index.


Leave a Reply