Transparency: the honest track record
Factual, verifiable performance evidence for EV-IDEX predictions — backtested, paper-traded and live results kept strictly apart, every number with its sample size and uncertainty, losses included. This page is updated automatically and shows no customer data.
Backtested = simulated on past data (can flatter). Paper-traded = Model v2's sealed shadow ledger, never shown to customers. Live = the public v1 ledger, sealed before outcomes were known. Every accuracy figure carries its correct/eligible counts and 95% interval; below 30 graded calls it is labelled "below the reporting threshold", and nothing is called meaningful below 100.
Cost model: Profitability after costs = signed return − 0.3% round trip (0.10% fee + 0.05% slippage per side), 1 unit per call. Neutral-band grading unchanged. Updated Fri, 13 Nov 2026 18:28:41 GMT.
2024-04-30 → 2026-10-26 · 14 coins · chronological walk-forward · reserved segment from 2026-04-28
shadow — never shown to customers. 113 calls still pending.
319 unique predictions (every record sealed) · 149 pending · 170 finished · 138 graded · 52 directional graded · 88 independent graded · 241 historical duplicates (audit only, never scored) · 171 superseded
Proof-of-concept status
Two different questions, answered separately: does the machine work (data in → sealed prediction → automatic grading) and are the predictions any good. Everything below is computed live from the public ledger; nothing is estimated.
A model only shows skill if it beats the best of these over many calls. In a one-way market "always Bearish" or "always Bullish" can look brilliant by luck.
One unit per call: buy on Bullish, sell short on Bearish, close at the horizon. Costs are charged once per round trip. Spot-only users cannot short, so the long-only row is the realistic one for most people.
Independent sample: one call per asset, timeframe and horizon window; Bullish/Bearish market calls only; neutral-band outcomes excluded from the denominator; baselines graded with the same ±2% rule on the same windows.
Live evidence ledger — what is reportable today
Each row is a real, sealed sample. A tier describes how much the sample can say — it never certifies accuracy.
| Sample | Type | Graded | Accuracy (95% CI) | Avg net after costs | Buy & hold, same windows | Worst call | Tier |
|---|---|---|---|---|---|---|---|
| v1 live engine · 7-day directional calls | live ledger | 6 / 30 | not reported (< 30) | — | — | -10.1% | insufficient |
| v1 live engine · 1-day directional calls | live ledger | 0 / 30 | not reported (< 30) | — | — | — | insufficient |
| v2 shadow · 7-day BUY/SELL calls | paper / shadow | 38 / 100 | 57.9% (42.2–72.1) | -0.679% (-2.932–1.574) | -3.253% | -19.46% | reportable |
| v2 shadow · 7-day BUY-only subset (long-only study) | paper / shadow | 12 / 30 | not reported (< 30) | — | — | -19.46% | insufficient |
| v2 shadow · 7-day calls with stated confidence (calibration study) | paper / shadow | 38 / 100 | 57.9% (42.2–72.1) | -0.679% (-2.932–1.574) | -3.253% | -19.46% | reportable |
| v2 shadow · 24-hour BUY/SELL calls | paper / shadow | 99 / 100 | 57.6% (47.7–66.8) | -0.231% (-0.758–0.297) | -1.241% | -9.56% | reportable |
Identical to the Monday e-mail digest: counts are read from the sealed ledger for the last 7 days — nothing is estimated or back-filled. The Prediction Accuracy Score is published only once both thresholds below are met.
Every day at 05:00 UTC the recorder seals one AI signal and one rule-engine decision per covered coin (origin scheduled), using the same model, prompt and grading rules as user-triggered analyses. Retried runs only fill gaps — never duplicates. Credits are the market-data and AI calls each run spent.
| Run (your time) | Status | Universe | Records created | Duplicates | Skips | Credits used | Errors |
|---|---|---|---|---|---|---|---|
| Nov 13, 05:02 AM UTC cron | done | 13 coins | 13 · 11 AI / 2 rule | 47 | 5 | 7 market-data · 13 AI | 0 |
| Nov 12, 05:01 AM UTC cron | done | 13 coins | 18 · 18 AI / 0 rule | 42 | 5 | 1 market-data · 13 AI | 0 |
| Nov 11, 05:01 AM UTC cron | done | 13 coins | 8 · 8 AI / 0 rule | 52 | 5 | 10 market-data · 13 AI | 0 |
| Nov 10, 05:01 AM UTC cron | done | 13 coins | 13 · 13 AI / 0 rule | 47 | 5 | 8 market-data · 13 AI | 0 |
| Nov 9, 05:01 AM UTC cron | done | 13 coins | 9 · 8 AI / 1 rule | 47 | 6 | 5 market-data · 12 AI | 0 |
| Nov 8, 05:01 AM UTC cron | done | 13 coins | 13 · 12 AI / 1 rule | 47 | 5 | 1 market-data · 13 AI | 0 |
| Nov 7, 05:01 AM UTC cron | done | 13 coins | 12 · 11 AI / 1 rule | 48 | 5 | 10 market-data · 13 AI | 0 |
| Nov 6, 05:01 AM UTC cron | done | 13 coins | 23 · 22 AI / 1 rule | 38 | 4 | 8 market-data · 13 AI | 0 |
| Nov 5, 05:03 AM UTC cron | done | 13 coins | 22 · 21 AI / 1 rule | 38 | 5 | 10 market-data · 13 AI | 0 |
| Nov 4, 05:02 AM UTC cron | done | 13 coins | 46 · 39 AI / 7 rule | 14 | 5 | 10 market-data · 13 AI | 0 |
| Coin | Skipped | AI / rule | Most common reason |
|---|---|---|---|
| FIGR_HELOC | 10× | 0 / 10 | no Coinbase USD market or not enough daily candles for the rule engine |
| TRX | 10× | 0 / 10 | no Coinbase USD market or not enough daily candles for the rule engine |
| WBT | 10× | 0 / 10 | no Coinbase USD market or not enough daily candles for the rule engine |
| XMR | 10× | 0 / 10 | no Coinbase USD market or not enough daily candles for the rule engine |
| LEO | 9× | 0 / 9 | no Coinbase USD market or not enough daily candles for the rule engine |
| figure-heloc | 1× | 1 / 0 | market data delayed or empty — not recorded |
Retry no-ops ("already recorded today") are not counted here — a retried run only fills gaps and never duplicates. Coins without a Coinbase USD market get an AI record only; the rule engine needs exchange candles.
Coverage = coins in the latest v1 scanner run. Rule-engine decisions additionally require a Coinbase USD market with ≥ 260 daily candles. Coins without a Coinbase USD market therefore receive AI records only.
| # | Coin | Exchange market | Records per run | Scheduled records sealed |
|---|---|---|---|---|
| 4 | BNB · BNB | Coinbase USD | AI + rule | 19 |
| 1 | BTC · Bitcoin | Coinbase USD | AI + rule | 9 |
| 12 | DOGE · Dogecoin | Coinbase USD | AI + rule | 14 |
| 2 | ETH · Ethereum | Coinbase USD | AI + rule | 11 |
| 9 | FIGR_HELOC · Figure Heloc | no Coinbase market | AI only | 14 |
| 10 | HYPE · Hyperliquid | Coinbase USD | AI + rule | 13 |
| 15 | LEO · LEO Token | no Coinbase market | AI only | 15 |
| 13 | XMR · Monero | no Coinbase market | AI only | 13 |
| 6 | XRP · XRP | Coinbase USD | AI + rule | 13 |
| 7 | SOL · Solana | Coinbase USD | AI + rule | 11 |
| 8 | TRX · TRON | no Coinbase market | AI only | 14 |
| 16 | WBT · WhiteBIT Coin | no Coinbase market | AI only | 13 |
| 11 | ZEC · Zcash | Coinbase USD | AI + rule | 13 |
Methodology
- Live ledger (v1): every public prediction is written to a hash-chained ledger when it is made, before the outcome is known, and graded once at its horizon by fixed rules (neutral band for small moves). Any free account can verify the chain under Track Record.
- Shadow ledger (v2): Model v2 (deterministic quant engine) runs alongside v1 and records sealed predictions nobody sees. It is compared against v1 on identical dates; it is not used for any customer recommendation.
- Backtest: strictly chronological walk-forward over the full coin universe — indicators use only data available at each decision point, every model is scored on the same windows with the same costs, and a reserved final 20% of the period was never used to choose anything.
- Baselines: coin flip (50%) for accuracy, cash (0%) and buy & hold on the same windows for net return. A model only matters if it beats these after costs with an interval that excludes them.
- Statistics: Wilson 95% intervals for proportions, bootstrap intervals (2000 resamples) for mean net, two-proportion z-tests and paired bootstraps for model comparisons, Brier/ECE for confidence calibration.
Risk disclosures
- EV-IDEX is a research and paper-trading platform. Nothing on this page is investment advice, an offer, or a solicitation. No real trades are executed by EV-IDEX.
- Backtested results are simulations on past data with assumed fills and fixed 0.30% round-trip costs (0.10% fee + 0.05% slippage per side). They were produced after the fact and can look better than live trading would have.
- Paper-traded (shadow) and live ledger results are real, sealed, time-stamped predictions graded once by fixed rules — but they are still not real-money results: no capital was at risk, no market impact, no missed fills.
- Small samples carry wide uncertainty. We show the 95% Wilson interval next to every accuracy figure. Below 30 graded calls a percentage is shown only together with its correct/eligible counts and is labelled 'below the reporting threshold'; even 100 calls do not prove an edge unless the interval excludes the baseline.
- Directional accuracy ignores calls that ended inside the neutral band; sign-after-cost and average net per call count every call and are the honest profitability view. Losses, worst calls and drawdowns are shown deliberately.
- Cryptocurrency prices are highly volatile. You can lose all the money you invest. Past performance — live, paper or simulated — never guarantees future results.
Want the full statistical detail (walk-forward tables, regime tests, confidence calibration, long-only study)? Create a free account and open Prediction quality tab. Nothing on this page is investment advice.
