Exploratory Research

Toward Asimov's Psychohistory

What happens when you feed 10,000 years of civilizational data to a machine learning model? It finds patterns—but not laws.

0Civilizations
0.00Model AUC (CV mean)
0+Years of History

The Seshat Global History Databank is an international research project that systematically codes historical and archaeological data—administrative hierarchy, military technology, religious practices—across hundreds of societies spanning 10,000 years. I trained a Random Forest classifier to find patterns in civilizational duration, though duration is an imperfect proxy for stability.

Important: This is exploratory analysis, not confirmatory prediction. Cross-validation shows AUC ~0.67 with high variance (0.51-0.76). Temporal holdout (LOEO) drops to 0.57—the model learns era-specific patterns, not universal laws. I'm working toward more rigorous methods.

What the Model Learned

Five findings from training a Random Forest on 256 polities across 10,000 years

01

Complexity alone explains nothing

The Tainter hypothesis needed context to work

0.505AUC (random chance)

I started with Joseph Tainter's classic argument: complex societies should be more fragile. The first model using only complexity features hit 0.505 AUC — literally a coin flip. Complexity matters, but only in combination with era and other factors.

+ More
02

The complexity curse reversed over time

What killed Ancient polities helped Early Modern ones survive

-159 → +6coefficient shift

In the Ancient world (pre-500 BCE), each unit of hierarchy reduced expected duration by ~159 years. By the Early Modern period, the relationship flipped — complexity slightly helped. Writing, institutions, trade networks changed the rules.

+ More
03

Religion outweighs military

Religious variables account for 27% of model decisions

27%feature importance

Religious institutionalization shows stabilizing effects. Ideology scores matter for fine-grained distinctions. Moralizing religion shows mixed effects — possibly reflecting rigidity or schism risk. The relationship is nonlinear: more isn't always better.

+ More
04

Warfare unlocked the signal

Adding military features improved classification by 28%

+28%AUC improvement

The model went from coin-flip (0.505) to meaningful signal (0.648 AUC) when warfare variables were added. Cavalry, armor, and fortifications moderate how complexity associates with duration.

+ More
05

The Classical sweet spot

500 BCE – 500 CE shows unique dynamics across all analyses

+0.634moderation effect

Rome, Han China, Persia — the Classical era consistently emerges as exceptional. Warfare moderation peaks here (+0.634 effect). Complex societies with strong militaries outlasted their simpler neighbors. Something about that combination, in that moment, worked.

+ More

Want the full methodology and analysis?

Or ask questions directly

Try Research Assistant BotBeta