TAIL RISK

CALIBRATION

OPEN SOURCE

Applied mathematics · systems · evidence

DresdenGoehner

2026

Building decision systems that make uncertainty visible.

Scroll to explore

Research direction

Independent work · 2023—2026

Models are useful only when we know where they fail.

I work across probabilistic modeling, energy systems, machine learning, open-source science, and products for real-world decisions.

The throughline is simple: replace impressive claims with calibrated evidence, explicit provenance, and tests designed to falsify the model.

Enter the complete archive →

Selected artifacts

Click through to the work

GERT public interface — GERT · live demo↗

EPSILON simulator interface — EPSILON · live demo↗

FutureOS guided prototype — FutureOS · local prototype↗

Images are working artifacts, not decorative stock photography. Each frame points to a repository, demo, or clearly labeled local prototype.

Evidence snapshot

Captured August 2026

65,711hours of continuous grid data

15merged upstream pull requests

22Math StackExchange answers

11accepted answers

Dated, independently captured records. No private documents are published on this site.

The model missed its calibration gate. That result mattered more than a polished demo.

Research note · GERT out-of-time evaluation

Flagship work

Systems · research · products

Work designed to survive contact with evidence.

Extreme systems

GERT interface showing grid risk signals — GERT · P50—P99 risk view↗

Grid Extreme Risk Toolkit · GERT

Seeing the tail before it becomes the headline.

Probabilistic grid-risk research built around P50–P99 forecasts, capacity margin, scenario intervention, provenance, and explicit simulated/live/stale labels.

65,711 hours of continuous grid data
Two strict out-of-time evaluations
Artifact and calibration gates
Public research and presentation modes

View repository ↗

PJM Polar-Vortex Study

Annual accuracy hid rare-event failure.

Fixed forecast origins and cold-event windows showed why ordinary validation can look acceptable while interval coverage collapses during the event that matters.

Read the study ↗

Decision laboratories

EPSILON

A trading simulator that asks which assumption breaks first.

A quantitative decision lab for testing orders, slippage, costs, drawdown, frequency structure, and falsification rules without presenting synthetic paths as historical validation.

Python, Next.js, and FastAPI generations
Deterministic synthetic public demo
Explicit non-investment and provenance labels

Open the live demo ↗

The Backtest That Lied

Leakage turned 0.94 into 0.10.

An audit of 200+ LightGBM fits that rejected an apparently exceptional strategy after correcting leakage, costs, and liquidity assumptions.

Open the audit ↗

Products with memory

FutureOS

Decisions should remember why they changed.

A local guided prototype for Decision Inbox, Decision Contracts, Living Future Maps, and Decision Memory—designed to retain reasoning as probabilities and plans evolve.

250 recorded tests
16 routes
Five-scene guided demo
Local prototype, not a public production deployment

AI Navigation Glasses

Assistive vision with safety boundaries.

A low-cost multimodal navigation system exploring traffic signals, obstacles, voice feedback, privacy, and maintainability for visually impaired users.

View the concept repo ↗

Complete work index

19 records · filter by practice

Not a highlight reel. A working archive.

Projects are grouped by the question they ask, not only by the technology used. Public links are included where available; prototypes and supporting records are labeled honestly.

Energy · probabilistic systems

Grid Extreme Risk Toolkit (GERT)

A decision-facing toolkit for tail-risk forecasting: P50–P99 intervals, capacity margin, scenario intervention, provenance, and strict simulated/live/stale labels.

Public repository · active buildOpen record ↗

Energy · research paper

PJM Polar-Vortex Study

A cold-event evaluation that tests fixed forecast origins, interval coverage, and why annual averages can conceal failure exactly when the grid is stressed.

Research artifactOpen record ↗

Quantitative decision lab

EPSILON

A trading simulator for orders, slippage, costs, drawdown, frequency structure, and falsification rules, with synthetic demo paths clearly marked.

Public repository · live demoOpen record ↗

ML audit · finance

The Backtest That Lied

An audit of 200+ LightGBM fits: leakage, liquidity assumptions, and costs turned an apparently exceptional 0.94 result into a defensible 0.10.

Public research projectOpen record ↗

Product prototype · decision memory

FutureOS

Decision Inbox, Decision Contracts, Living Future Maps, and Decision Memory—so a changed probability keeps the reason it changed.

Local prototype · 250 tests · 16 routesEvidence on file

Assistive AI · safety

AI Navigation Glasses

A low-cost multimodal navigation concept exploring traffic signals, obstacles, voice feedback, privacy, and maintainability for visually impaired users.

Prototype / design researchOpen record ↗

AI · information theory

NEE / Neuro-Entropic Engine

An exploratory architecture connecting entropy, information flow, and model behavior; the visual artifact is a research map, not a finished product claim.

Exploratory researchOpen record ↗

Mathematics · computation

Collatz Descent Framework

Computational experiments around descent structure and stopping-time questions for the Collatz map, with conjectures separated from verified observations.

Public research notebookOpen record ↗

Mathematical modeling

HiMCM Pareto Analysis

A competition modeling workflow using Pareto frontiers and trade-off analysis to make multi-objective recommendations legible.

Successful Participant · supporting artifactEvidence on file

Mechanics · bilingual explanation

AP Physics Tennis-Racket Theorem

A visual treatment of rigid-body rotation, polhodes, and the intermediate-axis theorem written to make a difficult physical idea teachable.

Research / teaching artifactEvidence on file

Data analysis · consumer behavior

Sneaker Market Intelligence

Product, price, platform, and demand analysis turned into a reproducible market-reading workflow.

Portfolio analysisEvidence on file

Education · project record

High-school PBL Portfolio

A structured archive of problem-based learning work: question design, data collection, group process, and reflection.

Public portfolioOpen record ↗

Research · pricing psychology

IPQ Charm Pricing

An independent project on charm pricing, experimental framing, and how behavioral claims should be separated from interpretation.

Research projectOpen record ↗

Statistics · visualization

30-project Data Lab

Thirty applied analyses across consumer behavior, finance, housing, sentiment, public health, environment, transport, and labor.

13 individual · 17 group projectsEvidence on file

Self-directed learning

SAT Practice Systems

A set of practice, error-log, and review systems for turning missed questions into deliberate iteration rather than a single score.

Personal learning systemsEvidence on file

Community · founder / president

Polymath Society

A student-led interdisciplinary society connecting mathematics, science, economics, and creative explanation through talks and project work.

Leadership record · brochureEvidence on file

Product · community operations

Club Platform

A platform concept for clubs to publish work, onboard members, and preserve institutional memory beyond a single school year.

Product conceptEvidence on file

Operations · software concept

Nexus Club OS

A club-operations system exploring structured events, member records, and repeatable workflows. Not presented as a production deployment.

Public concept repo · not deployedOpen record ↗

Industry · applied research

Modern Physics / Energy Innovation

Bilingual technical communication and a summer data-strategy internship that connect abstract models to energy and innovation decisions.

Academic + industry recordEvidence on file

Research gallery

Images from the evidence library

Polhode diagram from AP Physics work — AP Physics · polhode↗

HiMCM Pareto frontier analysis — HiMCM · Pareto frontier↗

Neuro-entropic engine research map — NEE · entropy map↗

Bike sharing feature importance chart — Data lab · feature importance↗

NetworkX contribution proof — Open source · NetworkX↗

Polymath Society brochure — Leadership · Polymath Society↗

Data laboratory

30 applied analyses

Many datasets. One habit: make the question explicit.

These projects move from collection to cleaning, visualization, modeling, and interpretation. The list keeps group work visible while distinguishing individual work from shared production.

13 individual projects

Tmall coupon behavior
Double 11 beauty conversion
Uniqlo customer segmentation
Xiaohongshu spending patterns
Domestic vs international box office
Spring Festival box office
Credit-card fraud detection
PPDai credit risk
Second-hand house scraping
Olympics performance
Airline customer value
Bike-sharing demand
Hotel booking cancellation

Individual source files remain in the private evidence library.

17 group projects

Seasonal vegetable prices
Advertising effectiveness
JD sentiment
Maternal / infant products
Supermarket baskets
Postgraduate score visualization
Weibo sentiment
50 Pyecharts examples
Water-quality image analysis
House-price modeling
Second-hand house market
Census age structure
National census
Marriage status
Tourist attractions
Data analyst jobs
Employee turnover

Open public group index ↗

Open-source engineering

15 independently verified merged PRs

NetworkX pull request evidence — NetworkX #8735 · merged upstream↗

Execution, audited in public

Contributing where the scientific details are sharpest.

These are code changes accepted into upstream projects—not stars claimed for the pull requests. Each item links directly to its public review record.

01argo-cd

Documentation / workflow quality

↗02FastStream

Framework correctness

↗03BeeWare

GitHub Actions security

↗04RepoWise

Repository tooling

↗05PEFT

Parameter-efficient training

↗06Dify

Developer experience

↗07Fairlearn

MetricFrame confidence intervals

↗08Fairlearn

Public API plotting

↗09SALib

Sobol analysis edge cases

↗10scikit-lego

Estimator behavior

↗11NetworkX

PageRank convergence docs

Conformal intervals

Typing / API quality

Time-frequency loss

AUROC / AUARC uncertainty metrics

↗

Mathematical practice

Public record · dated snapshot

978Math StackExchange reputation

22answers

11accepted answers

37answer score sum

Explaining is part of the proof.

Alongside competitions and formal certificates, public answers show how I reason in the open: definitions first, edge cases next, and a willingness to correct the story when the evidence disagrees.

View public profile ↗

Competitions & honors

Primary evidence + supporting records

BMO Round 2 Distinction certificate — BMO R2 2026 · Distinction / Best in School↗

LSE Top Gold certificate — LSE 2025 · Top Gold↗

BMO R2 2026

Distinction · Best in School

BMO R1 2025

Distinction

record

SMC 2025

Gold

LSE 2025

Top Gold · participation

AIME

Qualified · participation record

record

AMC 12

Top 10 record

record

Stanford Math Tournament

Algebra 3rd · Discrete 7th

record

HiMCM

Successful Participant

WSC

Official two-page team record

record

ACSL

Participation record

record

Certificate images are shown only for public-safe achievement evidence. Private transcripts, raw score reports, email attachments, and identity-sensitive records stay in the local evidence library.

Profile in one frame

How the pieces connect

Curiosity becomes useful when it can be shared.

I build at the seam between mathematics, technology, and people: a model needs a decision, a product needs a reason, and a team needs a memory.

That is why the archive includes code, diagrams, competitions, data work, leadership, and the limitations between them.

Industry

Energy Innovation · data strategy intern

Leadership

Polymath Society · founder / president
Club platform · community systems

Learning

Linear algebra · trading & machine learning · market microstructure

Polymath Society · leadership record↗

Evidence, not decoration

Public-safe index

Every claim should lead somewhere.

Public repositories, dated API captures, methods, limitations, and original project records form the verification layer behind this site. The local evidence library remains the source of truth.

DresdenGoehner

Models are useful only when we know where they fail.

Work designed to survive contact with evidence.

Seeing the tail before it becomes the headline.

Annual accuracy hid rare-event failure.

A trading simulator that asks which assumption breaks first.

Leakage turned 0.94 into 0.10.

Decisions should remember why they changed.

Assistive vision with safety boundaries.

Not a highlight reel. A working archive.

Grid Extreme Risk Toolkit (GERT)

PJM Polar-Vortex Study

EPSILON

The Backtest That Lied

FutureOS

AI Navigation Glasses

NEE / Neuro-Entropic Engine

Collatz Descent Framework

HiMCM Pareto Analysis

AP Physics Tennis-Racket Theorem

Sneaker Market Intelligence

High-school PBL Portfolio

IPQ Charm Pricing

30-project Data Lab

SAT Practice Systems

Polymath Society

Club Platform

Nexus Club OS

Modern Physics / Energy Innovation

Many datasets. One habit: make the question explicit.

13 individual projects

17 group projects

Contributing where the scientific details are sharpest.

Explaining is part of the proof.

BMO R2 2026

BMO R1 2025

SMC 2025

LSE 2025

AIME

AMC 12

Stanford Math Tournament

HiMCM

WSC

ACSL

Curiosity becomes useful when it can be shared.

Industry

Leadership

Learning

Every claim should lead somewhere.

GitHub

Math StackExchange

Work archive

Evidence library