Chapter Four · failure evidence
What Stochastic Modeling got wrong, from 59 dissertations
The records document various failures and rejections encountered when implementing stochastic models, optimization frameworks, and probabilistic algorithms. Researchers frequently found that stochastic formulations were outmatched by simpler deterministic baselines, hindered by severe computational demands, or undermined by high sampling variance and theoretical inconsistencies. These records come from PhD theses at 25 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Stochastic models frequently fail to outperform deterministic baselines
Across multiple application areas, stochastic planning, forecasting, and optimization models performed no better than or worse than deterministic baselines. In addition, increasing stochasticity in neural networks or policy optimization degraded predictive accuracy and objective values relative to deterministic counterparts.
Tried and failed
stochastic occupant behavior modeling applied to building energy performance simulation. Outcome: worse than baseline. Reason: failed to consistently outperform simplified deterministic models across simulation studies
Semantic Data Modeling for Quasi-Sentient Systems: A Framework for Holistic Knowledge-Based Automation · Georgia Tech
Tried and failed
two-stage stochastic programming applied to multi-sector energy system planning under climate uncertainty. Outcome: worse than baseline. Reason: value of the stochastic solution was negligible compared to expected value deterministic planning
Multi-system optimization : intermittent production, flexible demand, emerging technologies · UT Austin
Tried and failed
stochastic dynamic policy optimization applied to spatial dynamic closure decisions. Outcome: worse than baseline. Reason: training under stochastic conditions yielded no performance gain over deterministic baseline policies
Spatial and Behavioral Models for Ecological Management · Cornell
Lost to a baseline
Deterministic interdependence modeling (CS5: €1,776/day) outperformed stochastic ARIMA uncertainty modeling without interdependence (CS2: €1,737.94/day).
Integrating Low-Carbon Technologies in Electricity Markets: Network-Constrained Bidding Strategies for Distribution System Aggregators · Research Repository UCD
Considered and rejected
Considered and rejected: Rejected stochastic checkpoint strategies after proving theoretically that only a deterministic advancement boundary (Dirac delta function in time) maximizes arithmetic mean offspring.
Considered and rejected
Considered and rejected: Rejected modeling stochastic wave spectrum / probability distribution functions in favor of deterministic linear plane progressive wave theory.
Lost to a baseline
Stochastic Volatility (SV) models were beaten by deterministic GARCH-type models (FIGARCH/CGARCH/EGARCH) in out-of-sample forecasting for >60% of regions.
Lost to a baseline
On the gap splits of UCI regression benchmarks, increasing the percentage of sampled stochastic parameters under HMC degraded predictive performance relative to MAP and lower-stochasticity models.
Lost to a baseline
Fully stochastic Wide-ResNet-28-10 trained with MFVI achieved lower accuracy (94.69% vs 95.61% deterministic on CIFAR-10; 77.68% vs 79.33% deterministic on CIFAR-100).
Tried and failed
Sample Average Approximation with Benders decomposition applied to stochastic vehicle routing assignment. Outcome: worse than baseline. Reason: failed to outperform deterministic benchmark within practical runtime budget
Lost to a baseline
During the wet first LEMA period (2013-2017), the stochastic RHP model achieved lower contribution margins than the deterministic model configured with long-term average precipitation.
Integrating Groundwater Conservation and Risk Mitigation under Uncertainty: Strategies for Sustainable Aquifers and Agriculture · Virginia Tech
High dimensional state spaces and scenario explosion cause computational intractability
Explicitly calculating stochastic transition matrices, full posterior distributions, and multi-particle fields often proved computationally prohibitive. Researchers routinely abandoned full stochastic formulations in favor of deterministic or simplified approximations to avoid explosive scenario scaling and combinatorial complexity.
Considered and rejected
Considered and rejected: Rejected stochastic/method-of-moments precipitation modelling in favor of a deterministic law-of-mass-action kinetic framework to keep coupled multi-physics analysis tractable.
An Integrative Approach to Dynamic Processes in Reverse Osmosis Desalination · University of Nottingham Repository
Considered and rejected
Considered and rejected: Flattened stochasticity in topological evolution (zeroth-order dynamical flow) instead of full SDE, to produce a minimal, non-cumbersome viable model.
Evolutionary dynamics, topological disease structures, and genetic machine learning · OpenBU
Considered and rejected
Considered and rejected: Rejected exact computation of the posterior predictive distribution for iterative predictions due to analytical intractability over stochastic inputs.
Probabilistic reach-avoid for Bayesian neural networks · Oxford
Tried and failed
Direct time-domain simulation of fluctuating stochastic dipole ensembles applied to 3D far-field scintillation emission modeling. Outcome: infeasible cost. Reason: Simulating large ensembles of stochastic 3D dipoles in the time domain is computationally intractable.
Light-Matter Interactions with Photonic Quasiparticles · MIT
Tried and failed
treating recovery parameters as stochastic variables applied to uncertainty set computation. Outcome: infeasible cost. Reason: introduced combinatorial complexity and nonlinearities into computation without meaningful accuracy gains
Quantifying the Unknown: Data-Driven Approaches and Applications in Energy Systems · EPFL
Tried and failed
finite difference method for partial differential equations applied to high dimensional stochastic optimal control. Outcome: infeasible cost. Reason: computational complexity scales exponentially with state dimension beyond low dimensional systems
Advancing frontiers of path integral theory for stochastic optimal control · UT Austin
Considered and rejected
Considered and rejected: Rejected full 3D steady-state stochastic analytical formulation (Appendix E) due to excessive required mathematical approximations, relying instead on numerical simulation.
Stability and reversibility of compressible fluid displacement in porous media · UT Austin
Considered and rejected
Considered and rejected: Rejected transported PDF methods solved via stochastic Monte-Carlo particles for practical engine simulations due to prohibitive computational expense.
Conditional source-term estimation evaluations for partially-premixed flames · Oxford
Considered and rejected
Considered and rejected: Rejected pure stochastic programming due to scenario explosion and scarce future renewable adoption data in favor of possibilistic-robust programming.
Optimal Operation and Planning of Multi-Energy Microgrids with Hydrogen Carriers Integration · Osuva
Considered and rejected
Considered and rejected: Rejected calculation of the explicit state transition probability matrix P for the EV routing MDP due to stochastic driver behavior, dynamic traffic conditions, and unmodeled charging station wait times.
The digitalization of energy systems: towards higher energy efficiency · EPFL
Stochastic formulations struggle with structural misrepresentation and unidentifiable parameters
Pure stochastic models often failed to capture underlying non-random objectives, non-linear pricing trends, or empirical performance distributions. In several instances, stochastic estimation was abandoned because model parameters were unidentifiable or generated distorted probability distributions.
Tried and failed
purely stochastic binary-choice drift diffusion model applied to baseline behavioral choice performance distributions. Outcome: did not generalise. Reason: fails to capture empirical spread of baseline performance without incorporating attentional state transitions
Decision Making, Behavioral Development and Fine Motor Control In Larval Zebrafish · Harvard
Considered and rejected
Considered and rejected: Rejected stochastic sampling/dropout of Bayesian feature selection masks during test-time inference in favor of deterministic thresholding to obtain consistent subject diagnosis labels
Lost to a baseline
An MA(50) moving-average stochastic volatility model produced an unnaturally narrow 10-year FIA return distribution (SD of 0.006) similar to a constant-volatility GBM, failing to match historical FIA return dispersion compared to rough fractional Brownian motion models.
Selected Computational Problems In Insurance · YorkSpace
Considered and rejected
Considered and rejected: Rejected using pure stochastic Markovian/econometric models alone because they fail to exploit non-linear deterministic trends in historical pricing
Techno-economic analysis and method development applied to an aerobic gas fermentation and supercritical water gasification process · University of Nottingham Repository
Considered and rejected
Considered and rejected: Purely stochastic WaveFunctionCollapse unit-block generation was rejected as an optimizer due to its inability to navigate non-random objectives effectively
Generative constructal design for thermal flow systems · Imperial
Considered and rejected
Considered and rejected: Decided against using stochastic simulation models in favor of numerical differential equation solving because stochastic models cannot easily track local spatial concentrations required for 2nd-order bimolecular reactions.
Dynamische Ionenaustauschchromatographie reaktiver Aluminium-Komplexe Dynamic ion exchange chromatography of reactive aluminium complexes · open_UMR Marburg DSpace 10.0
Considered and rejected
Considered and rejected: Rejected using annual consumption data because state-space models with stochastic volatility are poorly identified at annual frequency.
Essays on Asset Pricing and International Finance · ResearchWorks
Considered and rejected
Considered and rejected: Rejected the topographic elevation (η-based) classification for calibrating stochastic vegetation models because it yielded non-unimodal, irregular pdfs that degraded statistical moment evaluation compared to inundation probability (PI-based) classification.
Integrated approaches for monitoring and modeling vegetation in riparian and coastal environments · IRIS - POLITO - prod
Tried and failed
stochastic process modeling of decision confidence applied to human perceptual discrimination datasets. Reason: model parameters were unidentifiable, leading to extremely low model recovery probabilities
How do humans give confidence? Comparing popular process models of confidence generation · Georgia Tech
Considered and rejected
Considered and rejected: Assigning stochastic CdV data points directly to deterministically trained HMM states was rejected due to mean state drift leaving regimes unpopulated.
On the variability and forced response of atmospheric regime systems · Oxford
Excessive noise and high variance disrupt stochastic estimation and sampling
Stochastic sampling techniques such as importance sampling and backpropagation through sampled variables suffered from high variance and inaccurate gradient estimates. Excessive stochastic noise also hindered optimization convergence, obscured experimental comparisons, and triggered unstable phase divergences.
Tried and failed
decision-focused learning with gradient backpropagation through sampling applied to network security game optimization. Outcome: worse than baseline. Reason: inaccurate gradient estimates from backpropagating through randomly sampled stochastic variables
Optimizing Decision-Making under Uncertainty -- A Data-Driven Perspective · Georgia Tech
Tried and failed
Monte Carlo scenario generation with rolling volatility applied to stochastic asset allocation models. Outcome: unstable. Reason: scenario generation failed uniformity tests under true empirical volatility without artificial parameter dampening
Improving time series and cross-sectional momentum trading strategies using stochastic programming · Iowa State
Lost to a baseline
Uniform sampling in stochastic gradient descent outperformed importance sampling after 5 iterations (1364 samples) because importance sampling suffered from high variance.
Decision-making for autonomous agents in adversarial or information-scarce settings · UT Austin
Considered and rejected
Considered and rejected: Rejected training stochastic differential equations by integrating dynamics and optimizing over full probability distributions or multi-Gaussian fits due to failure under significant inherent noise.
Modeling Dynamics of Classical and Quantum Systems Using Machine Learning Techniques · Publikationssystem UB Tuebingen
Tried and failed
relying on numerical noise for startup applied to autonomous oscillator simulation. Reason: stochastic initial conditions caused phase divergence between deterministic and perturbed simulation runs
Jitter performance in high speed oscillators · Iowa State
Lost to a baseline
Robbins-Monro Stochastic Approximation sieve baseline converged substantially slower than the proposed UQ gradient and accelerated gradient methods on the benchmark quadratic function under Monte Carlo noise.
Considered and rejected
Considered and rejected: Rejected adding stochastic error components to machine-learned motion planning models because it exacerbates traffic stop-and-go oscillations
Longitudinal Control for Self-driving Cars with Traffic Flow Considerations: Theory, Design, and Experiments · Georgia Tech
Considered and rejected
Considered and rejected: Rejected simulating historical catastrophic high-severity fire regimes (e.g., 1868 fire) because of high stochastic variance obscuring comparisons and uncertainty in post-harvest fuel loads.
Theoretical inconsistencies and mathematical rule violations break stochastic formulations
Multiple methods failed when standard ordinary differential equation rules or operator equalities were incorrectly applied to stochastic processes. Other formulations broke down because history dependence destroyed detailed balance or because risk measures violated dynamic consistency across stages.
Tried and failed
static coherent risk measures applied to multistage stochastic optimization. Reason: they fail dynamic consistency and recursivity across multiple decision stages
Some Unconventional Stochastic Programs · Georgia Tech
Tried and failed
exact expectation conditioned on full state history applied to stochastic learning algorithms with Markov states. Outcome: infeasible cost. Reason: exponential growth of terms when conditioning on full historical state trajectories
The algorithmic learning equations · UT Austin
Tried and failed
equating functional derivative operators across stochastic calculi applied to infinite-dimensional martingale expansions. Reason: operators diverge without strict conditional expectation or adaptedness regularity constraints
Stochastic taylor expansions for functionals of martingales · Imperial
Considered and rejected
Considered and rejected: Stateful variance-reduced stochastic gradient estimates (e.g. SVRG) within standard discrete random walks, rejected because history dependence destroys detailed balance / exactness of the stationary distribution.
Efficient Sampling Using Markov Chain Monte Carlo Methods · ResearchWorks
Tried and failed
treating stochastic differential equations as ordinary differential equations applied to likelihood ratio computation with random coefficients. Reason: stochastic integrals with random coefficients do not follow standard ordinary differential equation calculus rules
Optimal parameter adaptive estimation of stochastic processes · Virginia Tech
Tried and failed
stochastic selective dissipation modeling applied to stochastic Lie-Poisson dynamical systems. Reason: anisotropic noise and mismatched dissipation prevent forming an exact Gibbs invariant measure without strict nilpotency or symmetric friction
On the rough diffusive limits of deterministic geometric mechanics · Imperial
Tried and failed
Stochastic reduced variational principles on semidirect products applied to ideal fluid dynamics. Reason: Hamiltonian is not hyper-regular with respect to the advected density parameter
Stochastic geometric mechanics for geophysical fluid dynamics and wave-current interactions · Imperial
Numerical discretization and bounding errors degrade stochastic solutions
Truncation horizons and implicit numerical integration schemes dampened essential fluctuations or produced bounds that contradicted deterministic benchmarks. Furthermore, strict probability thresholds and multidimensional collocation errors caused numerical infeasibility and poor convergence.
Tried and failed
single-value opportunity cost calculation applied to multi-metric stochastic portfolio evaluation. Outcome: did not generalise. Reason: scalar opportunity costs fail when evaluating multi-metric stochastic returns across heterogeneous scenarios
System of Systems Stakeholder Planning in a Multi-Stakeholder, Multi-Objective, and Uncertain Environment · Georgia Tech
Tried and failed
chance-constrained stochastic programming with deterministic approximation applied to distributed energy resource planning. Reason: setting an overly strict probability threshold rendered the optimization problem mathematically infeasible
Stochastic Programming Models for Planning Wind Based Distributed Generation in Prosumers of Energy Mode · TXST Digital Repository
Tried and failed
finite horizon simulation for infinite horizon bounds applied to discounted stochastic optimization problems. Reason: truncation horizon was insufficient for high discount factor, producing statistical upper bounds lower than deterministic lower bounds
Risk neutral and risk averse stochastic optimization · Georgia Tech
Tried and failed
implicit Euler-Maruyama numerical integration applied to stochastic branching and extinction dynamics. Reason: Implicit time-stepping artificially dampened stochastic fluctuations, severely underestimating variance and missing discrete extinction events.
Mathematical and computational models of nuclear reactor start-up physics and operations · Imperial
Lost to a baseline
For high stochastic dimensions ($N=56$, $L_E/l=0.1$), sparse SC method reference error failed to compete with standard brute-force Monte Carlo simulation.
Numerical treatment of imprecise random fields in non-linear solid mechanics · Leibniz Universität Hannover Repository
Left open by the authors
Problems the authors named and did not get to.
Left open
Rigorously prove that informative scheduling policies stochastically outperform non-informative counterparts across general queueing settings. Blocker: None
Information Freshness Optimization in Real-time Network Applications · Virginia Tech
Left open
Simulate routine screening effectiveness across various pathogen epidemiologic characteristics using the discrete-time stochastic network model. Blocker: None
Infectious Disease Surveillance and Control for Pandemic Prevention, Preparedness, and Response · Harvard
Left open
Derive analytical expressions for heteroplasmy temporal dynamics and mutant fixation probability across two or more coupled demes using the stochastic model. Blocker: None
Stochastic modelling and inference for evolution in ageing and infectious diseases · Imperial
Left open
Reformulate the compartmental epidemic model to incorporate parameter stochasticity using a Markov process instead of deterministic difference equations. Blocker: None
Strategies for Effective Mitigation of Infectious Diseases, with Focus on COVID-19 · Virginia Tech
Left open
Formulate robust geometric programming and chance-constrained optimization models for epidemic control on networks with stochastic transitions. Blocker: None
Left open
Fit realistic seasonal transmission and recovery rate functions to empirical epidemiological datasets for stochastic epidemic models. Blocker: None
Left open
Formulate and solve two-stage or multi-stage stochastic programming models for prescriptive epidemiological resource allocation. Blocker: None
Large-scale Optimization for Robust Multi-Class Prediction and Resource Allocation · MIT
Left open
Simulate stochastic survival of the densest models on two- and three-dimensional spatial deme structures. Blocker: None
Stochastic modelling and inference for evolution in ageing and infectious diseases · Imperial
Left open
Fit the stochastic spatial SEIR-SEI metapopulation model against multi-country arbovirus case series and simulate global transmission dynamics beyond single-run approximations. Blocker: None
Modelling climate-driven spatiotemporal transmission dynamics of aedes-borne arboviruses · Imperial
Left open
Extend stochastic extinction and outbreak probability calculations to epidemic models with mass action incidence or large natural demographic turnover. Blocker: None
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.