Daily Summary for 2026-10-15
daily
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: It's the fifteenth of October, twenty twenty-six, and this is the day's research.
Warren: One new paper came out today.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We'll take the day in one pass, then pull out the paper we're staying with.
Warren: Start us off with Mixture-of-Experts for Cryptocurrency Order Execution.
The day's research: Ruth: Mixture-of-experts models for crypto order execution stability is key.
Warren: How does that relate to tail risk in real trading systems?
Iris: We looked at training stability using vanilla Double Deep Q-Learning policies.
Ruth: Specifically, we evaluated K means partitioned mixtures of these experts.
Warren: What budgets did you test for those expert partitions?
Iris: We tested budgets of two, four, eight, and dense networks matched to four and eight.
Ruth: We used five-minute mean-aggregated BTC/USDT limit order book data from Binance.
Warren: Did the results show any improvement in mean implementation shortfall over vanilla DDQL?
Iris: No learned configuration significantly improved it.
Ruth: In fact, all tested methods had a higher mean shortfall than TWAP at zero point three nine basis points.
Warren: And immediate liquidation was measured at zero point two one basis points for comparison.
Iris: Under the reported specification, most vanilla-DDQL runs converged to a policy that waited until forced liquidation.
Ruth: That is good because it means the system did not crash in many cases.
Ruth: Annealing exploration alone eliminated observed collapses in twelve out of one hundred runs.
Warren: That matches the effect seen when using expert partitioning, correct?
Iris: Combining annealed exploration with aligned reward restored collapse in nineteen out of thirty runs.
Ruth: And with all three changes combined, it rose to forty-eight out of one hundred.
Warren: This suggests expert partitioning might be masking a training specification failure.
Iris: No mixture-of-experts run with K equals eight collapsed under any six tested specifications.
Ruth: The dispersion across seeds was lowest at K equals eight, but it was inconsistent.
Warren: It did not hold up when adjusted for family-wise effects, which is a key point.
Iris: Within a single policy, tail risk got worse as K increased.
Ruth: We have Mixture-of-Experts for Cryptocurrency Order Execution: Training Stability, Tail Risk, and Failure Modes because it shows that mixing experts does not improve performance in this specific trading task and that stability issues are related to training setup rather than the expert structure itself.
Warren: That confirms the finding regarding stability versus architecture.
Iris: Indeed, the focus shifts to training setup rather than just expert structure.
Ruth: It is important that we note this distinction for future work.
Lucky paper: 2610.03369: Ruth: So we're looking at how this Mixture-of-Experts for Cryptocurrency Order Execution paper explains the gap between execution cost and realized returns.
Warren: The paper sets up the problem by defining implementation shortfall as that gap caused by commissions, limited book liquidity, and reactive pricing.
Iris: It shows that vanilla Double Deep Q-Learning policies don't significantly improve mean implementation shortfall compared to those baseline models.
Ruth: That's a big finding since we often look for architectural gains when we try to fix execution problems.
Warren: The authors test different expert budgets, specifically K equals two through eight and dense networks for four and eight experts.
Iris: They found that across the seed dispersion, the lowest dispersion was at K equals eight, but it wasn't increasing monotonically across those seeds.
Ruth: And that relationship wasn't robust when they adjusted for family effects on the expert budgets.
Warren: Then they also observed that within-policy tail risk got worse steadily as they increased the number of experts used.
Iris: But a key result is that no Mixture-of-Experts run with K equals eight collapsed under any of the six specifications they tested.
Ruth: That suggests the apparent attribution of failure modes might just be an artifact of how many seeds you run.
Warren: The authors investigated specification decomposition, testing a reward that differs from the evaluation metric, and a non-annealed exploration schedule ending at epsilon equals zero point eight eight.
Iris: When they combined annealed exploration with the aligned reward, the collapse rate went up to forty-eight out of one hundred runs.
Ruth: That strong interaction between those three specification changes is what really reveals how sensitive these policies are to setup choices.
Warren: The paper reports that on identical episodes and fill mechanics, TWAP beats every learned configuration on mean implementation shortfall at zero point three nine basis points.
Iris: And the tail risk for TWAP is a CVaR-ninety-five of seventy-four point zero, which is better than every agent tested in that comparison.
Ruth: It’s striking that the RL execution policy with K equals eight got down to zero point zero one basis points for implementation shortfall, which is essentially execution at the arrival price because the replay buffer was frictionless.
Warren: That really highlights how much the environment itself matters when you look at these numbers from this Mixture-of-Experts for Cryptocurrency Order Execution paper.
Iris: The limitations mentioned are important because they show where this research stops working, like the historical replay being frictionless and not capturing nonlinear microstructure impacts.
Ruth: So even though we see these architectural changes reshaping the distribution of training outcomes, the training specification is what actually determines if any of them clear a schedule baseline.
Warren: It’s a diagnostic tool for understanding policy robustness rather than just another piece of execution software.
More episodes
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- The day in conversation for 2026-11-18
- Daily Summary for 2026-11-18
- 2601.05274-On the use of case estimate and transactional payment data in neural networks for individual loss reserving
- 2009.00868-Loss-Given-Default Modeling by Post-Last Passage Time Process
- 2606.30070-Financial Resilience Evaluation: From Conditional Expectations to Dynamic Convex Risk Measures
- 2212.12687-Measuring price impact and information content of trades in a time-varying setting
- Daily Summary for 2026-11-17
- 2401.08323-Dynamic portfolio selection under generalized disappointment aversion