Methodology
Every model begins as a claim that can be disproved. The research moves from historical replay to unseen-period tests and then to a forward paper study, comparing forecasts with recorded results at each stage. AI assistance is tested as a separate, costed treatment and never substitutes for the checks required before a Coinbase order.
Independent Evidence Programs
Round 2 paper and live trading may observe the same market, but they are not one experiment ledger. The paper run uses fixed virtual capital. A live deployment begins only when its Coinbase units, cash, prices, and benchmark units reconcile and are locked. This prevents a paper starting value from drifting with the real account and prevents a live deposit, withdrawal, or fill from being mistaken for model return.
Monthly Outperformance Standard
The objective is repeatable after-cost excess, not a guarantee that every calendar month will be positive. Crypto regimes change and any strategy can underperform. A model may be described as a consistent monthly outperformer only after it clears the following untouched-window contract.
Academic Evaluation Rubric
A model earns credibility one stage at a time: a clear theory, historical replay, tests on unseen periods, a forward paper result, comparison with an AI treatment where applicable, realistic cost and execution assumptions, and a new version only when the evidence explains why a change is needed.
Software And AI Pipeline Rationale
The research also depends on how evidence is recorded. Stable model names, clearly separated data sources, and dated snapshots make it possible to reproduce a result. AI suggestions are recorded separately from any permission to preview or place an order.
Finance Principles Under Test
The principles below explain why each comparison or safeguard is included. Use them to interpret the charts, evaluate a strategy definition, and decide whether gathering a non-committing Coinbase preview would add useful evidence.
Hypothesis Testing Routes
Different research questions require different views. The routes below show where to compare model families, critique one model, examine historical or unseen-period tests, review paper and AI results, inspect asset routes, or check preparation for a live test.
Inspect A Model Definition
Hypothesis:
Signal inputs:
Decision rule:
Risk rule:
Projection method:
Evaluation Standard
- Actual paper outcomes begin on
; current runtime evidence is through . - Production remains blocked until the Day 30 report and human confirmation gates are complete.
AI Assist And A/B Testing
AI Assist is tested as an addition to the rule-based strategy, not as a replacement for it. The study runs the unchanged model and an AI-assisted paper version over the same period, assigns the cost of AI review to the treatment, and asks whether the remaining improvement is large enough to justify another test.
Follow the AI question through three stages. Methodology defines the test. AI Assist Research compares the completed benefit with the cost and identifies which models, if any, merit another paper test. Strategy Setup lets a signed-in user apply that research decision to a draft strategy.
AI Experiment Implementation
Downside-Capture Treatment
A good explanation after a decline earns no AI lift. The alert must exist before the verified event start, state a horizon and probability, and use only contemporaneous evidence. The experiment then isolates prediction, defensive action, and re-entry so a profitable outcome cannot hide a bad forecast or a costly false alarm.
How AI lift is scored and discussed
- Predictive lift is the improvement in a proper probability score, such as Brier score, versus the unchanged parent forecast. Directional storytelling is not enough.
- Downside-capture lift is the risk-gate paper value at the verified trough minus the parent value at that same timestamp, after spread, slippage, and fees.
- False-defense cost is the value surrendered when the gate reduces exposure and the feared decline does not occur.
- Re-entry lift is the staged-re-entry ending value minus the risk-gate-only ending value. It prevents “sold before a decline” from being counted as success when the model never restores exposure.
- Net portfolio lift compares the complete alert, gate, and re-entry path with the unchanged control after trading costs and AI review cost.
AI Treatment Planning Estimates
The calculations below are planning examples, not experiment results. They estimate whether a proposed AI treatment is worth testing. If its estimated improvement does not cover the review cost, the proposal is rejected before spending. The completed cost-benefit result appears in AI Assist Research; the current round's recorded control and treatment appear in Strategy Tournament.
Methodology estimate
Estimated paper treatment arms
The model-only control starts at zero lift. The AI-adjusted point shows estimated net lift after review cost. Negative values mean that AI mode should not spend until recorded evidence improves.
Decision Policy And Live Gates
Decision policy is a model-selection filter, not an order ticket. Allowed assets, cash reserve, trade cadence, and drawdown limits express risk constraints that should narrow the candidate set before Strategy Setup. They do not authorize live Coinbase orders, and they are not personalized financial advice.
Model Families
How Predictions Are Produced
Where The Models Live
The models are mathematically defined as deterministic JavaScript rules and parameters. They are not black-box LLM prompts, and the AI helper layer is a separate paper/shadow experiment that can suggest parameter changes only after cost and risk gates pass.
Research Lineage
These are research-inspired model hypotheses, not faithful reproductions of any one published paper or a promise of performance. The app turns common investment concepts into explicit, testable rules so their behavior can be compared against Coinbase market data and the 30-day paper tournament.
Support, Resistance, and Market-Level Evidence
The model does not assume a named price is a wall. Articles and social posts may supply a timestamped hypothesis. Academic evidence determines how it is operationalized: as an expiring probability zone, discovered from pre-decision observations and tested against failed touches, random zones, simple rules, and the unchanged parent.
Source-by-source critical review
The strongest finding is about behavior and market structure: round prices attract orders, repeated zones may temporarily alter bounce probability, and near-touch liquidity can matter at short horizons. The weakest claim is that today’s $61K or $67K headline level will reliably reverse price. That fixed-level claim is not admitted without a timestamped definition and an untouched after-cost test.
From chart annotation to reproducible feature
We can test the supplied analysis without pretending the drawing is a result. The screenshot is recorded as
The unlabeled bottom oscillator in the screenshot is deliberately excluded. We will not infer an indicator or parameter set from appearance alone. This technical-structure branch remains paper-only until it produces incremental after-cost excess across untouched, non-overlapping months and survives indicator ablations.
Cross-Horizon and External-Context Research
The useful lesson is a hierarchy, not a larger indicator pile. Completed daily evidence may define a regime, completed hourly evidence may define an entry, a separately tested exit may manage the position, and forecast volatility may adjust paper risk. Macro, blockchain, treasury, and vendor observations enter only at their first public timestamp and must add lift beyond simpler price-only controls.
Candidate sequence
Each row is a separate test. The execution challenger cannot borrow credit from the multi-timeframe alpha parent, the treasury branch cannot treat vendor estimates as filings, and the external vendor stream remains a benchmark rather than becoming an opaque ensemble vote.
OANDA’s historical narrative is used only to motivate structural-break stress tests. The actual regime labels are computed from completed price returns and realized volatility using thresholds frozen inside training. Post-event stories never become model inputs.
External Research Integration
External forecasting projects offer useful techniques—including neural networks, time-series models, volatility models, sentiment features, and diagnostic workflows—but adopting a technique changes the research hypothesis. New methods should therefore enter as versioned offline experiments, while the existing rule-based control and its recorded paper history remain unchanged.
Current model-family branch
For the tree-ensemble paper, the app adopts only the low-cost challenger hypothesis. It replaces price-level R² with return/direction diagnostics and after-cost parent/benchmark excess, requires naïve forecast controls, and fits preprocessing inside each training fold. Lean contributes the separation of forecast, portfolio, risk, and execution contracts—not a new runtime dependency.
The historical replay currently uses the public Coinbase candles API as its source of OHLCV market history. The app can be made more precise by adding minute-level candle windows for intraday strategies, but the current design deliberately avoids expensive full tick-history storage while we are still testing whether the concept is worth deeper infrastructure.
Runtime Cadence And Signal Horizons
Model evaluation separates signal calculation from trade execution. The low-cost runtime can refresh public market data every