The question of how we can learn useful lessons from history, which would help us navigate the troubled waters ahead has been much on my mind. You can find some of my thoughts on this subject in my review of Walter Scheidel’s Escape from Rome. It’s also an important theme in the “Big Book” I am working on (there will be an update on it soon). And then yesterday I saw a Twitter thread by Balaji Srinivasan in which he asks the same question. Balaji mentions my work and that of Ray Dalio, but is somewhat critical of us. He says:
The main knock on Turchin & Dalio is that their quantification is imprecise. What’s the y-axis here? Seems like a weighted sum of many variables.
and gives an example from Dalio’s site. This critique may be correct when directed at Dalio (as I remember I similarly tried to get at his data sources some time ago, but couldn’t find it). But this charge is certainly not true when directed at my work. It is all explained in Ages of Discord, which is an academic book and has references for all data sources that I used in it (and that’s a lot). I also started a webpage on which all these data will eventually be posted, then had to set this project aside for a while. Now my research assistant is working on not only posting the data, but extending them to the present. Stay tuned.
To reiterate: all data on which my graphs are based are sourced in the academic pubs and the methods used for generating various curves from these data, such as the one below, are properly described and can be reproduced by anybody. When the webpage I mentioned is fully operational, this will be even easier.
As to the substance of Balaji’s proposal, I obviously agree with the basic idea that we need to construct and analyze historical databases, since that is what I’ve been doing since 2010 (see here the latest example). But there are good reasons to believe that the specific approach that he proposes for analyzing historical data will not work. What he describes is an atheoretical, entirely data-based analysis. Such a “brute-force” approach has the following problems:
1. There are thousands of variables that characterize societal dynamics. In the absence of theory, how do we choose the relevant ones? If we don’t select, the curse of dimensionality will defeat any attempt at trajectory-matching, which is essentially what he proposes.
2. Trajectories of specific societies are constantly perturbed by exogenous shocks (climate, external wars, epidemics…) and endogenous sources of variation (free-willed people behaving in unpredictable ways). Also, because social system are characterized by complex nonlinear feedbacks, they operate in a chaotic regime (in technical terms they are characterized by sensitive dependence on initial conditions). Thus, the trajectories of two identical dynamical systems starting from identical initial conditions will rapidly diverge (because sensitive dependence will amplify even small shocks within each system) and lose any resemblance of each other. This will defeat any attempt at trajectory matching.
So what can we do? First, stop treating societies as black boxes. We have pretty good ideas about the mechanisms that generate social dynamics. So we need to use explicit mechanism-based models. In fact, the problem is the opposite: there are too many ideas about what are the important mechanisms. Thus, the first use of a historical database is to empirically test different mechanistic theories against each other. Next, we can use the data to estimate such things as the signal/noise ratio and fine-tune parameter estimates (initial estimates should come from other sources than time-series as much as possible).
I can write more about this, but it’s too big a theme for a blog post, so I’ll stop here. As a final thought, the way I envision the eventual product, coming from this work, is a multi-path forecasting engine, the prototype of which is described in this publication.
I want to emphasize that these critical comments are “tactical” in nature. I completely agree with Balaji on the need to construct comprehensive historical databases that avoid the problems of cherry picking and the bed of Procrustes. In fact, our CrisisDB project is doing something very similar to what he wrote in the tweets, collecting time-series data on economic, political, social, and cultural variables of societies as they slide into a crisis, and then emerge from it. Perhaps there is a potential for collaboration here. In any case, the hard part is collecting the data, once there is a good database, different teams can (and will) use different analytical approaches to extracting insights from the data.