Fig 1.
CCS Jaffe + Sectors (α = 0.1).
In the CCS (Continuous Company Space) representations above each point represent a company. On the upper panel, points are colored according to companies’ industrial sectors. The effects of the information on the industry are visible from the clustering of firms of the same color. On the lower panel, we represent on the CCS some acquisitions performed by three large companies as arrows from the acquirer firms to the targets. This is an example of how acquisitions are likely to be done locally in this space.
Fig 2.
Predictions evaluation of Target and Pair forecasts.
Comparison between the prediction performances of the different similarity metrics. We investigated different values of class imbalance, that is for each true M&A we extract 1, 20, and 200 negative cases (random couples with no deals). In both figures, we use three different colors to distinguish the typologies of metrics. Blue represents direct metrics, yellow indirect ones, and orange is the CCS ones. The error bars are computed by repeating the extraction of random Target companies and/or Acquirer companies 20 times. A: Comparison among the Best F1 scores on Target Prediction (fixed acquirer). B: Comparison among the Best F1 scores on Pair Prediction (both target and acquirer are not fixed). In both cases, Best-F1 values are normalised to ones of random predictions. With this normalisation, a Best-F1 value around 1 will indicate a random-like prediction, while values greater than 1 will correspond to more significant predictions. The Jaffe + Sectors metrics outperforms all other approaches.
Table 1.
Best-F1 values table of Fig 2.
Jaffe + Sectors (J + S) constantly outperforms all the other measures on both Target and Pair Prediction and for all class imbalances. On the other hand, Euclidean Distance (EU) is always the worst performing. Due to the normalisation with respect to the random case, the Best-F1 values increase with the class imbalance, but the ranking of the different measures remains nearly the same.
Fig 3.
Dependence of maximum Best F1 on α and the class imbalance in the Jaffe + Sectors measure.
For each value of class imbalance, we rescale the Best F1 between 0 and 1 to better visualize the maximum as a function of α. The maximum F1 moves towards higher α values when the class imbalance increases. This suggests that when choosing a M&A pair among a large pool of options, the industrial sector plays a more important role.
Fig 4.
Behavior of Best F1 versus α for various class imbalance values in the Jaffe + sectors measure.
Due to the correlation between Best F1 and class imbalance the various curves never intersect. In this plot, three phases divided by dotted lines can be spotted: a Low α phase, at the bottom-left, where there exists an α-dependent maximum for the Best F1, a High α—Low-class imbalance phase, at the top-right, and the High α—High-class imbalance phase at the bottom-right, where the Best F1 is independent of α.