Backtests

A backtest asks whether a model's historical result survives different starting dates rather than depending on one fortunate entry. These replays use daily Coinbase open, high, low, close, and volume data. Models designed for intraday trading are approximated with each day's high-low range, so their results should not be read as minute-by-minute fill simulations.
Round development record

What historical evidence changed the frozen model set?

Backtests screen timing sensitivity before a round begins. They inform selection, but they are not appended to the later paper return and they never turn a retrospective estimate into a locked forecast.

Selection evidence

Multi-start stability and unseen-window return

Each point is a model frozen into . Rightward means a higher hit rate across historical starts; upward means a higher average return across untouched walk-forward windows. Color identifies whether it was added, retained, or an initial candidate.

Boundary: this view explains why a candidate entered the round. It does not prove that the candidate succeeded in that round. Continue to Walk-Forward for the round forecast and actual comparison.

1. Choose Models

Select one or more model rows to cross-filter replay outcomes. The compact table keeps repeated metrics aligned for comparison and uses the full 30-day replay; the controls below then change horizon, entry type, regime, date range, and error tolerance.

2. Cross-Filter Historical Starts

Where The Test Positions Were Set

The charts below restore the complete position map to the default workspace. The large BTC chart shows the common historical test clock. Each smaller chart then shows the sampled starts derived independently from that asset's own Coinbase price history. The date, entry-regime, and market-regime controls above filter every position marker; model and error filters affect replay outcomes, not the shared sampling clock.
Common test clock

BTC price history and filtered test positions

The line is the Coinbase daily close. Each dot is a replay start, colored by the condition that placed it in the test set. Hover to inspect its date, price, entry class, and market regime.

3. Asset Hypothesis Lens

This lens asks two related questions: which historical training points define the selected asset class, and whether model evidence changes when filtered to windows that include that asset. Existing replay rows remain BTC-clocked for continuity with past snapshots; the asset toggle below now also exposes each product's own OHLCV-derived training/start candidates for asset-specific model review. Exposure summaries are not causal proof; they show how windows using an asset behaved.
Asset training points

price history and sampled model starts

The line is the selected asset's Coinbase daily close. The dots are the sampled start points derived from that same asset's 90-day range, drawdown, trailing return, and volatility. Use this to inspect the asset-specific training candidates before interpreting strategy outcomes.

Asset exposure outcomes

Realized return by asset exposure

Each bar summarizes filtered strategy windows that included the asset in their modeled allocation. This helps identify whether model evidence is broad across assets or concentrated in one exposure.

4. Asset Lens

Use this table to follow one model, asset, and holding period from the inputs it observed to the selection it made and the return that followed. These are historical replay records; they are kept separate from the later paper study.
Parameter derivation

Selection metric by sampled start

The vertical axis is the model-specific selection metric: momentum for momentum models, risk-adjusted momentum for volatility-adjusted models, drawdown pressure for reversion models, and range-volume pressure for swing harvesters.

Prediction evidence

Predicted return versus realized return

Selected dots show the evidence that affected the modeled sleeve. Rejected dots remain visible so the user can see whether the model passed over assets that later performed better.

5. Predictions Versus Realized Outcomes

Calibration

Predicted return compared with realized forward return

Points above the diagonal beat the model's estimated return; points below it underperformed. This is the main view for judging whether a model's historical prediction method was directionally useful.

Entry-date behavior

Realized return by start date

This view shows how sensitive the selected models are to when the strategy begins. Use the date brush above to focus on specific market eras without computing every possible historical start.

6. Compare Horizons

Horizon sensitivity

Average realized return by forward horizon

Short horizons are useful for monitoring, while longer horizons show whether a model survives broader market drift. This chart keeps your model, date, entry, and regime filters but compares all replay horizons.

How To Read This

Sample clockBTC-USD defines the market clock because it has the longest Coinbase daily candle history and drives broad crypto regime changes.
Entry labelsPeak, dip, middle, and rolling starts are derived from BTC's local 90-day range and drawdown from the 90-day high, not hand-picked narrative dates.
Forward horizonA 30-day horizon means the model is switched on at the sampled start date and scored against the next 30 daily closes.
Model replayThe replay uses daily-close allocations, cash reserves, model-specific asset choices, estimated fees, and drawdown adjustments. It is intentionally lighter than a tick-level execution simulator.

Boundaries

This is a backward-looking replay layer, not a promise that the same conditions will recur. It uses true historical Coinbase candle data for start points and forward returns, but the model behavior is an approximation of each strategy's rules. It does not model intraday fills, order-book depth, taxes, outages, exact spread at execution time, or exchange-specific order rejection.

Legacy Score Model

The final table preserves the score model that preceded the historical replay generator. Treat it as a record of the earlier ranking method; use the replay charts above when evaluating whether a prediction matched the later historical outcome.
Prior ranking

Legacy 90-day model score

This is useful for continuity with the earlier app, but it is no longer the main explanation of historical sampling.