The nonlinearity in a source of stress is necessarily associated with fragility. This is perhaps why low-quality coins fail --marketing activity is an environmental variable only, while actual installed capacity and operational infrastructure (i.e bitcoin's) is a robustness contributor factor. Once we have established the driving variables, we need to study the volatility of those variables (and the non-linear relationships). We see how with high volatility the predictor performs poorly, at least with this small subset of variables, which are also all highly correlated. A possible fragility test would be to derive the "volatility of volatility" where the price volatility prediction becomes a traffic prediction problem. But the key realization, in alignment with Taleb and Douady, is that to be fragile the system has to be non-linear to harm (has to be accelerated to harm). Since fragility that comes from linearity is immediately visible, the hidden risks and potential harm come from non-linearity. In this example, the AI notes that the activity in a particular mining pool frequented by the users of the exchange has a low correlation to volatility, and will use the second derivative of volatility respect to the mining pool (usage per day is the event size) as a "trust" distance metric. The formula is obtained using symbolic regression as well, computing a smoothing spline for volatility with respect to mining pool, and then, computing the symbolic derivative of the spline functions (cubic polynomials), and evaluating the expression at various data points. One of the possible models has the form D(Volatility, (Mining pool), 2) = (sin(1.27652458974758 + 1.63551681837977e-5*(Mining pool)) + cos(3.2702109619178e-5*(Mining pool) + sin(2.14970699758616e-5*(Mining pool))))/(Mining pool)
Of course, the data set can be enhanced with multiple data sources, and the prediction error reduced by combining additional methods (e.g. de-noising with recurrent neural networks and convolutional neural networks). But by providing a context of the environment to the intelligent agent, the AI is in a better position to reason on the trustability of the result.
Spatio-temporal patterns in blockchain networks
Space of Production
In one of the first studies of the disciple of Human Geography into distributed shared ledgers, Blankenship \cite{blankenship2018} conceives blockchains as production spaces where developers are the dominant class within the social and technical spaces of the technology, have ultimately leveraged their knowledge and power dynamics to accumulate wealth via the token value, and then shifted into the role of investors. This necessarily involves automation (exploitation of automated robot labor) and obfuscation of the mechanisms of production -- geographic borders are defined via conflicting abstract conditions (social, political, and economic), and put within the qualitative context of social dynamics. Humans not only trust in the source, but they also trust the structure -- you generally do not care about who wrote a diet article (even if a change in lifestyle can have a lasting impact on health) as long as "structure" suggest the writer is not a charlatan. A similar behavior is observed in crypto markets, where traders and investors keep lists of Twitter accounts that they trust on to relay accurate information about the state of the market, and that is facilitated by other traders: it does not count only who is saying it, but who is following -- this is part of the social fabric of crypto markets, the structure of the network encodes tacit knowledge and reflects abstract conditions and boundaries. The distance trust metrics have very tangible implications for individual and corporate purposes; "a member of my group said " (even if he had materially different attributes) is generally better than what an outsider says. The implications in terms of the theory of the firm \cite{2014}: you do business in the proximity of your circle (your trust space) where trust is secured, even if it is more expensive to produce in your inner circle, and it is cheaper to acquire in the boundary (e.g. potential partners) -- but going beyond that will require a significative leap of faith and the associated risk should be priced-in. The AI will understand this human inclination (as shown in Figure 2), as a trust differential in terms of metric entropy.