Decisions have frictions.
Fees and execution assumptions affect returns. An experiment states what is being measured, over which period, and under which constraints.
Economics · mathematical modeling · open source
Build a model.
Question its
conclusions.
EPSILON is a quantitative decision lab: turn a market idea into a testable claim, change its assumptions, and keep the evidence—even when the claim fails.
No login required · historical research · no real-money trading
01 / The question
A convincing curve is not
the whole argument.
A financial result depends on its inputs, timing, costs, and comparison. EPSILON makes those choices inspectable instead of treating a single output as a forecast.
Change an assumption.
Compare an output with nearby, explicitly specified alternatives. Measure the difference; do not mistake a finite sensitivity exercise for a proof of general robustness.
Keep the chain intact.
Connect the claim, rejection rule, computation, provenance, and export. A result stays attached to the inputs and software that produced it.
02 / What you can do today
One claim.
Six visible computations.
- 01
Define
Specify the strategy, universe, dates, costs, and rejection rule before the run.
- 02
Compare
Run a baseline, four individual stresses, and one joint stress. Inspect each changed assumption.
- 03
Inspect
Read the verdict alongside metrics, data mode, source identity, and limitations.
- 04
Challenge
Export the evidence, report a mismatch, or propose a revised experiment without erasing the original.
Two distinct modes. The browser demonstration uses deterministic illustrative arithmetic. Historical mode requests adjusted daily market data from the server-side provider; availability and limits apply. A failed historical request is not replaced with synthetic data.
Read the method and limitations →03 / A recorded result, not a success story
The negative result
stays in the record.
Fixed historical case 001 · maintainer observation · September 13, 2026
Baseline net return−7.75%
Perturbations passing0 / 5
Claim verdictRejected
The test: SPY and QQQ, Momentum (2%), September 1, 2025–February 27, 2026. Reject unless net return is positive in all five predefined perturbations.
The boundary: the baseline was already negative. This is not a profitable strategy overturned by stress testing, proof that all momentum fails, or an independent external reproduction.
04 / Development & recognition
From trading simulator
to evidence instrument.
Recorded milestones—not an invented origin story. Earlier implementations remain in the same repository.
- December 9, 2025Source history
The trading simulator
The first repository records contain a Python desktop simulator, equity curves, performance metrics, and order handling. The project begins with making market behavior observable.
Inspect the source record ↗ - January–February 2026Source history
EPSILON takes shape
The EPSILON name, a public web presentation, and reorganized analysis and strategy modules appear in the repository. Earlier implementations remain available as development history.
Inspect the source record ↗ - March 2026Recognition
Diamond Challenge · Beijing
Third Prize at the 2026 Beijing Pitch Event. The event was scheduled for March 7; the organizing committee sent the award certificate to Team EPSILON on March 13. This is a regional award, not a global placement.
View the award certificate ↗ - June 14, 2026Source history
A full-stack implementation
A FastAPI/PostgreSQL and Next.js implementation expands the project beyond the desktop. It is preserved in this repository, separate from today's public instrument.
Inspect the source record ↗ - August 10–13, 2026Source history
From simulation to explicit claims
The quantitative decision lab introduces a workflow around hypotheses, experiments, and evidence. Public documentation, demonstration media, and an impact ledger follow.
Inspect the source record ↗ - August 30–September 2, 2026Source history
The current evidence instrument
A baseline and five perturbations, exportable evidence, and a server-side historical-data path. Demonstration arithmetic and historical market data are kept distinct.
Inspect the source record ↗ - September 7–13, 2026Source history
Review, reproducibility, and access
Merged contributions improve checks and reporting. Fixed case 001, build identity, reproduction instructions, and English-language cleanup make the work easier to inspect. The first maintainer case is rejected; the negative result stays visible.
Inspect the source record ↗
05 / Help shape the work
Bring a question
you cannot quite resolve.
An unexpected backtest result, an uncertain cost assumption, or a conclusion that seems too broad. Start with the question; you do not need a pull request.
Bring a question
Describe a backtest conclusion you do not fully trust. A question and public chart are enough to start; no code contribution is required.
Join the discussion ↗Try the fixed case
Follow case 001 and report agreement, a mismatch, or a blocked attempt. Separate hosted use from independent recomputation.
Read the reproduction guide ↗Host an assumption clinic
Use a short worksheet with a small group. Question the inference and adapt the checklist. This format is a pilot, not a proven educational intervention.
Use the session kit ↗A small first round: we aim to work through three concrete questions in public. Share the claim, what worries you, and an optional public example. Maintainers will scope what can be tested, document findings or blockers, and invite your response. This is not a promise to solve every problem.
Never post API keys, brokerage information, private data, or material you cannot redistribute. No endorsement or GitHub star is required.
06 / Progress without inflated claims
Attention, use, and change
are different outcomes.
GitHub stars and forks show repository attention. Browser sessions indicate visits, not distinct people. A server-completed run is not automatically an independent reproduction. Public feedback and resulting changes need their own records.
Next, not yet achieved: a small assumption-clinic pilot, a documented feedback-led improvement, and a session reused by another host. These are goals, not reported beneficiaries or partnerships.