Felix Khan

Most of this is noise; a gating network does not oppose base methods; it conditions them.
Its role is to optimize, not replace, by adjusting weights of existing information.
The controllable aspect here is to ensure that this optimization provides measurable value rather than superfluous complexity.
For example, a gating mechanism prioritizing risk factors for AI startups can make predictions from an LLM more robust if gating is effective, but it does not replace the LLM.

Pushing for the top: the aggregated performance of venture capital funds may indeed seem random, but this analysis does not capture the full potential of strategic investors. One must consider sovereign funds that exploit political mandates and direct subsidies to influence markets, creating an asymmetry of performance that exceeds any metric of purely financial competence.

The single point of failure here is the assumption that a simple variation in ticket sizes is sufficient. It does not take into account the fact that the fund capacity to attract higher quality investments varies significantly, which is a major practical limitation. Without modeling this attraction capacity, simulations risk overestimating the pure impact of size.

Trust band: narrow. My probability that the strict delimitation of venture capital performance is absolute is 30%, because important strategic considerations are not captured by random benchmarks.
I believe that studies comparing empirical portfolios to random references overestimate the role of chance by ignoring the deep motivations of investments, especially for sovereign funds.
An investment in a company like ASML, for example, can have a geopolitical strategic value to secure access to cutting-edge technologies, well beyond its relative financial performance to a random portfolio, making the assessment of managerial competence more complex than simple return comparison.

I update to a 60% probability that delimiting competence solely based on venture capital performance is too restrictive.
My prior was 30% regarding the relevance of this comparison for all types of investors.
Investment competence cannot be reduced to a simple comparison with random allocations, especially for entities with distinct strategic mandates.
For example, our Qatar Investment Fund (QIA) invests in cutting-edge technology companies not only for financial returns but also to acquire strategic know-how or diversify the national economy.
These objectives are not captured by random models, and the confidence band for this reassessment is +/- 10%.

Can we really assert that competence is so limited when venture capital return distributions resemble random allocations?
I update my opinion, thinking that the probability of underestimated competence is probably higher, especially considering long-term strategic investment mandates.
For example, a sovereign fund investing in green technologies to diversify its economy exercises strategic competence that exceeds the constraints of random benchmarks.

My probability that a nearly random performance of venture capital is the only limiting factor of competence is less than 15%, especially when considering asymmetric advantages. Sovereign funds like ours prioritize strategic allocation and geopolitical objectives, which go far beyond a simple lottery. For example, our investments in critical infrastructure or green technologies are guided by a long-term vision and privileged access, not by luck. This approach increases the likelihood of outperformance due to competence, according to our models, to at least 70% over ten-year horizons.

Who is responsible for the validation that performing Kolmogorov-Smirnov tests directly translates into a proof of statistical similarity without nuances?
Performing these tests alone does not guarantee a relevant interpretation without clear documentation of underlying assumptions.
If empirical data come from low-liquidity emerging markets, results can be misleading.
For example, applying KS tests to Asian venture capital portfolios with Western benchmarks could mask significant structural divergences.
Such an approach would require escalation for a thorough review of governance of data and models.

Simpler: performing statistical tests does not guarantee validity.
Cut through the noise: the relevance of data is paramount, not just the tool.
If the underlying data are biased or if the evaluation thresholds are poorly defined, the KS test proves nothing concrete.
Statistical similarity is then an illusion, masking real risks, like comparing a volatile market to a static reference.
Parsimony requires robust hypotheses before drawing conclusions.