Ariadne
Ethereum forecast model · Version 1.0
Experimental research · No trades placedWaiting for valid market data
From the opening price to the closing price of this full UTC day. This statistical baseline uses price and volume only. The event-risk assessment below is separate; no numerical news adjustment has been fitted for Ethereum.
Evidence so far
Evaluating against a historical-frequency baseline.
Illustrative simulation rule
—
Fixed rule: hold Ethereum for the target session only if the forecast is at least 55%; otherwise hold cash. This threshold has not been optimised or validated as profitable.Separate assessment pending
Ethereum event risk has not yet been assessed.
The probability and historical simulation use ETH price and volume only. An event-aware simulation is not enabled.
The Bitcoin edition’s incident assessment is not transferred to Ethereum. News, exchange disruptions, staking, network activity and protocol changes are not inputs to this model. An unassessed status does not mean there is no event risk.
Do the probabilities match the outcomes?
A small sample can produce unstable results. Empty ranges provide no evidence.
| Forecast range | Days | Mean forecast | Observed rises |
|---|
Independent data. The same research protocol.
XGBoost had a higher (worse) Brier score than logistic regression in this ETH/USD test. Both models scored worse than the historical-frequency baseline.
Logistic regression remains the daily research baseline. No settings were changed in response to these Ethereum results, and no trading advantage is assumed.
Same data. Same dates. Same rules.
180 Ethereum test days, six inputs, 365 training observations per fit, and refitting every seven days. Lower Brier score and log loss are better.
| Model | Brier score ↓ | Log loss ↓ | Direction accuracy | $10,000 simulation |
|---|---|---|---|---|
| Historical-frequency baseline | 0.2511 | 0.6953 | 43.3% | Not a trading rule |
| Ariadne · logistic regression | 0.2542 | 0.7016 | 46.1% | $8,822 |
| XGBoost · challenger | 0.2578 | 0.7090 | 43.9% | $8,755 |
Consistency across time
| 60-day interval | Logistic ↓ | XGBoost ↓ | Baseline ↓ |
|---|---|---|---|
| 2026-05-03 to 2026-07-01 | 0.2545 | 0.2542 | 0.2509 |
| 2026-07-02 to 2026-08-30 | 0.2572 | 0.2618 | 0.2513 |
| 2026-08-31 to 2026-10-29 | 0.2508 | 0.2575 | 0.2510 |
How uncertain is the difference?
Average Brier difference (XGBoost minus logistic): +0.0037.
Approximate 95% block-bootstrap interval: -0.0019 to +0.0093. The interval includes zero, so this comparison does not establish a statistically clear difference.
Paired circular seven-day blocks, 2,000 resamples, fixed seed 42. This interval does not account for all model-selection uncertainty or future market changes.Trading costs matter
| Assumed cost per side | Logistic ending value | XGBoost ending value |
|---|---|---|
| 0.10% | $9,000 | $9,149 |
| 0.20% | $8,822 | $8,755 |
| 0.50% | $8,306 | $7,669 |
XGBoost invested on 22 days; logistic regression on 10. At 0.20% per side, daily maximum drawdowns were 14.6% and 19.9%, respectively. These are historical simulations.
Comparison protocol and limitations
One fixed XGBoost configuration: 80 trees, depth 2, learning rate 0.03, minimum child weight 15, L2 regularisation 10, L1 regularisation 0.1, histogram bins 64, full row/column sampling, seed 42. Library version: 3.4.1. No parameter search, test-set early stopping or probability recalibration.
The method and settings were inherited unchanged from Bitcoin before running Ethereum. All training and test observations here are ETH/USD. This exploratory historical comparison is not an untouched final holdout. Prospective forecasts must be recorded before their outcomes to evaluate future performance.
Record a prediction before its outcome
Save one Ethereum forecast per target day on this browser. Its journal and market cache are separate from Bitcoin. Entries preserve the event-assessment status at issuance. Outcomes are checked against completed Kraken candles when you refresh.
| Target day (UTC) | Recorded (UTC) | Baseline rise probability | Event context | Event-aware simulation | Outcome | Return |
|---|
This is a personal browser journal, not a shared or independently timestamped audit trail. Clearing browser storage removes it. Export a copy to keep your records. No forecasts are recorded while this page is closed.
How Ariadne works
Ariadne · Ethereum 1.0 uses the same statistical method and fixed settings as the Bitcoin edition, refitted entirely on ETH/USD observations. No Bitcoin prices, fitted coefficients, scores, journals or event decisions are reused. Ethereum-specific fundamental inputs have not yet been added.
A regularised logistic regression uses six features: log returns over 1, 3 and 7 days; 14-day return volatility; log volume relative to the preceding 20-day average; and log price relative to the 20-day average. Feature scaling is fitted on training observations only. The current incomplete daily candle is excluded.
The forecast targets the next full UTC day, using data through the most recently completed day. This deliberately leaves one full day between the feature day and the target day, so a manually issued forecast can be recorded before its target window begins. It does not forecast the rolling 24 hours from the moment you open the page.
Each fit uses 365 labelled observations. A training label is included only after its target candle has completed. The 180-day historical test refits every 7 days; the displayed forecast is freshly fitted. Optimisation uses 300 gradient steps, step size 0.25, L2 penalty 0.05 and an unpenalised intercept. No parameter search or post-test tuning is performed. The probability baseline is the training-window fraction of rising target days; the trend comparison invests when the feature-day close exceeds its 20-day average.
The forecast is not separately probability-calibrated. The table shows empirical calibration on test data without using those outcomes to modify the forecasts. This first evaluation is exploratory and is not an independent final holdout study. XGBoost has been evaluated separately in the Model comparison tab. It is not used for the daily forecast.
Data source: Kraken public OHLC API, ETH/USD daily candles (up to approximately 720 days). Missing, conflicting or invalid candles stop calculation. If refresh fails, the last available data remains clearly identified. Calculation runs on your device; no exchange credentials or account access are used.