EtherForecastEtherForecast · Service status
Confirming service status and current forecasts.
Lights on: current forecasts available. Lights off: maintenance or status not confirmed. Forecast quality is reported separately below.
Animation could not load. The service status remains available above.
Explore Ethereum forecasts from 6 hours to 30 days. See probabilities, price ranges, and how published predictions compare with actual outcomes.
This window also applies to the performance below.
Update status and predictive performance are separate. Live errors use only settled, timely published forecasts. No-change assumes the reference price stays the same; lower MAE is better. Overlapping windows share price movements.
Historical simulation
Reconstructed tests on past data. Results below are for the selected forecast window, period, and market conditions.
Green: forecast midpoint. White: actual outcome. Shading: the model's intended 80% range. Prices are compared at the same end date.
Lines show the estimated chance of touching each threshold. Dots at 1 mark periods when that move happened. Both moves can happen within one period.
The forecast window and market filter stay the same. Positive improvement means less error than the past-frequency baseline. Overlapping forecasts are not independent trials.
Brier score measures probability error; lower is better. Alert thresholds target a 5% false-positive rate in earlier validation data. The rate on later data can be different. Catch rate counts detected events; false-positive rate counts alerts among periods without that event.
Models are selected monthly using outcomes already available then. Market conditions use trailing 24-hour returns and 30-day volatility. High volatility means at least 1.5 times the 30-day level. Only complete, settled, matching observations are scored. Historical receipt times are reconstructed, so these are not live results.
Loading test details.
Published forecast record
Only forecasts published before their observation window count here. This record uses the selected forecast window across all available live dates; the historical filters above do not apply. Pending outcomes are not counted as wins or losses.
All six windows · complete history
The forecast window above selects the detailed charts. These filters apply to every table and chart in this workspace, including the six-window comparison. Historical tests and published forecasts are scored separately.
Waiting for verified archive.
Published and experimental charts include pending forecasts with their original ranges. The actual-price line appears only after settlement. Pending forecasts and late or unverified publications do not enter the scores.
Points compare predicted probabilities with observed rates in ten bins. Small bins are uncertain; hover for sample counts and descriptive binomial intervals, which do not adjust for overlapping windows. Use the non-overlapping count when assessing evidence. The two path events can both happen.
MAE and RMSE measure forecast error; lower is better. Coverage compares the original 10th–90th percentile range with settled outcomes. Interval score penalizes both wide ranges and missed outcomes. Pinball loss uses the three published quantiles; it is not an exact CRPS.
A complete historical study has not been published yet.
Loading candidate review.
Only matching, timely published, settled windows with the same realized outcome are compared. Historical development scores above are separate.
The midpoint is the model's central estimate. Lower, within range, and higher probabilities describe where the final price lands relative to the two stated move thresholds. These three probabilities add to 100%.
A large-rise probability means the price touches the upper threshold at any point during the window. The price can rise and then fall, so the two touch probabilities do not have to add to 100%.
Less probability error does not mean the same percentage of predictions were correct. Check missed moves and false alarms too. An 80% price range is a model target; its actual hit rate is shown under Performance.
Observation starts at the next full UTC hour after the forecast's input slot. Every forecast shows its exact start and end. The live quote does not change an already published forecast. Model edges and live performance remain under evaluation.
Previous research, retained for comparison. The hourly event outlook above is the current forecast system.
CatBoost + patch Transformer
Complete chronological replay
Model settings and blend weights use only outcomes available before each monthly fit. Price inputs are closed daily bars; flow inputs have an extra day of delay. Historical source vintages are reconstructed. Retrospective rankings do not establish live predictive skill.
Download predictions and outcomes (CSV.gz)Earlier forecasts and evaluations are retained for comparison. These archived statuses do not describe the current hourly outlook.
Archive Checking archive Checking evaluation
Archived reference: — · —
Forecasts stay fixed after issue. A 30-day forecast needs 30 days before its outcome can be scored.
Loading daily research evidence…
These experimental forecasts are saved when issued and scored after their 7-day or 30-day target. UP / FLAT / DOWN probabilities cover all outcomes. Improvement over the no-change benchmark is not yet established.
MAE skill above zero means lower error than no change on the same dates. Formal review needs at least 90 resolved forecasts and six non-overlapping 30-day blocks. Research results do not automatically replace the official forecast.
Daily live evidence
Each point is added only after its target date has matured.
Live h=7
history loading
Live h=30
history loading
Historical research evidence
evaluation loading
Each monthly fit uses earlier data only. “Best full-period” is selected retrospectively from the eight tested candidates. “Past-only selection” chooses using outcomes already known at each fit. These are reconstructed historical tests; they do not replace the live forecasts above.
Historical vendor publication vintages remain unverified. Positive MAE skill means lower error than no change. Direction scores below cover UP / FLAT / DOWN outcomes and are not probability calibration.
Download all replay predictions (CSV.gz)OOF h=7
model loading
OOF h=30
model loading
Support
Voluntary support helps with data collection, hosting and model evaluation. Forecast access stays free; donations do not guarantee investment returns or forecast accuracy.
Choose the listed asset and network. Public wallet addresses only; balances and transaction totals are not shown.
Keeping a useful public forecast takes more than a model: it takes continuous data collection, reliable hosting and honest testing. Your support helps preserve the data available at each forecast time, evaluate new signals and keep the research open to everyone.
CoinGlass Standard + CoinGecko Analyst
This is a planning estimate, not a current bill or a requirement to buy both. We would test one source at a time before an annual commitment. Business quotes, tax, overages, storage and computing are additional costs.
Official public prices checked 12 September 2026 (UTC), in USD, with annual billing. These are proposed data sources; paid-data integration and additional forecast gains have not been demonstrated.
| Source | Annual cost | Data to evaluate | Why test it? |
|---|---|---|---|
| CoinGlass Standard ↗ | $3,588 / yearCommercial plan · $299/month equivalent | Cross-exchange open interest, funding rates and liquidation history; ETH ETF flows where entitled. Data catalogue ↗ | Test whether leverage and forced selling add warning signals at 6 hours–3 days. Standard hourly history reaches up to 360 days; minute data is shorter. Verify endpoint access before buying.Additional forecast gain: unmeasured. |
| CoinGecko Analyst ↗ | $1,238 / yearPublished annual total · attribution required | Aggregated prices, market capitalisation and trading volumes; up to 10 years of history where available. | Test whether broader market context and more complete historical coverage help the 1–30 day windows. Some price data is already available free; paying alone does not create a new signal.Additional forecast gain: unmeasured. |
| Santiment Sanbase / SanAPI ↗ | Business quote requiredIndividual Max reference: $2,700/year. Excluded from the total above. | Social activity and sentiment, on-chain wallet and exchange activity, subject to metric and ETH coverage. | Test attention spikes and network activity as context for 1–14 day forecasts. The individual price is not a confirmed public-site licence; real-time API, history and commercial terms must be checked.Additional forecast gain: unmeasured. |
| Glassnode Professional + API ↗ | Annual quote requiredAPI access, credits and usage rights must be included. | ETH on-chain flows, realised-value and staking metrics where supported. API plan details ↗ | Test whether supply and holder behaviour help the 7–30 day windows. The cheaper Advanced subscription is not equivalent to a full historical API licence.Additional forecast gain: unmeasured. |
We do not yet have a measured uplift from these subscriptions. More data can produce no improvement or even worse forecasts. It would be misleading to promise a percentage increase in accuracy before a matched test.
Proposed pilot hurdle: 3% lower price error or 5% lower probability error. These are evaluation targets, not expected gains. Before each pilot, choose one primary metric: price MAE or Brier score. Compare the added data with the same frozen model without it, and check that other metrics and forecast windows do not materially deteriorate.
Illustration only: Brier 0.20 → 0.19 is a 5% relative reduction in probability error; MAE $100 → $97 is a 3% error reduction. Neither means accuracy rises by 5 or 3 percentage points. These numbers are not EtherForecast results.
Report each of the six windows separately using time-ordered tests, data actually available at the issue time, gaps between training and test targets, and non-overlapping outcomes with uncertainty ranges. Long windows need much more history. An inconclusive or harmful result is a reason to defer adoption.
Your donation can fund licensed historical data, collection and storage, model evaluation and the infrastructure that keeps the site available. The aim is to make better-supported forecasts and show whether a change helped—not to hide unsuccessful tests. Data sources would be adopted in stages, with their coverage and measured impact reported alongside the research.
Forecast access stays free. Support is voluntary and does not purchase a trading signal, a promised accuracy level or an investment return.
후원은 유료 데이터 이용권, 과거 데이터 보관, 수집·서버 운영과 예측 검증 비용에 쓰일 수 있습니다. CoinGlass Standard와 CoinGecko Analyst의 후보 예산은 합계 연 $4,826이며, 현재 지출이나 필수 구독을 뜻하지 않습니다. Santiment와 Glassnode의 사이트용 이용 조건은 별도 견적이 필요합니다.
유료 데이터의 실제 예측력 향상은 아직 측정되지 않았습니다. 가격 오차 3% 또는 확률 오차 5% 상대 감소는 파일럿 평가 목표이며 예상 성과가 아닙니다. 한 가지 주 지표를 미리 정해 동일 조건에서 비교하고, 개선이 없거나 악화되면 도입을 미룹니다. 후원 여부와 관계없이 예측은 무료로 열람할 수 있습니다.
Wallet details are loading. If they remain unavailable, please check again later.
도움이 되었다면 무료 예측 운영을 후원해주세요. 한국어 이용 안내