our work
MIGO
ROBOR8
MIGO: machine learning via graph optimization
MIGO belongs to a class of Artificial Intelligence called Machine Learning (ML). It uses a semi-supervised learning algorithm where the system is given “samples” used for training. A sample is an [input,target] pair and the objective is to minimize the miss-classification of the targets.
MIGO uses a type of ML algorithm called Classification Trees, but here is where any similarity to any other tree-based algorithm ends. Compared to other traditional ML algorithms (such as Neural Networks or Support Vector Machines), classification trees have the unique property that, however complex the derivation algorithm may be, the solutions themselves are easily understandable and simple to explain, giving insight, sensitivity and therefore trust into the answer. Inputs to the classification tree are called “factors” and the set of inputs used to create a single tree is called the “factor-set”. For MIGO, the factor-sets are derived using a branch of mathematics called Graph Theory.
MIGO creates a collection of classification trees, called a “forest”. The outputs from each tree of the forest are then combined using an evolutionary algorithm, to produce the final output of the system.
The MIGO classification trees are created by proprietary, graph optimization algorithms which guarantee (with mathematical proof) that the produced trees are optimal. In addition, the generated trees possess another unique and superior property: namely, they incorporate both orthogonal and skewed cuts of the sample space. The trees also have a deep-learning characteristic which is unique to MIGO. The resulting classification trees have superior performance to any other existing system.
For white-papers or questions regarding the MIGO system, please email us.
The rating generator
roboR8 is a corporate credit rating system based upon a new approach in machine learning which combines deep learning with optimal classification trees.
The roboR8 methodology takes a set of financial ratios, together with macro-economic data, to create the system inputs. From these inputs, the model optimally chooses the subset which is most relevant and the chosen subset is then used as input into a machine-learning classifier which creates a non-linear “mapping” from the input to the output (the rating). The inputs have no reliance upon traded instruments and, therefore, the same methodology can be applied to public, private, large or small companies.
Our Machine Intelligence via Graph Optimization (MIGO) classifier that performs the above mapping is proprietary and is based on an optimal algorithm developed by our research team. The superiority of MIGO over other existing ML algorithms has been demonstrated both theoretically and in practice on this and several other applications.
The roboR8 system using the MIGO rating generator produces a Rating Distribution rather than a simple rating. The full rating probability distribution of possible ratings is produced which can be used to:
1. Assess the confidence of the assigned rating
2. Present more detailed risk insight by giving a probability of a lower/higher rating
3. Compare same-rated corporates: companies with the same single-label rating do not possess the same risk and the distributional rating output can be used to make an objective credit comparison between companies