Home › Guides

How accurate are AI trading signals?

Updated 2026-08-28 · 894 signals tracked

• 60.7%Win rate
• 1.27Avg R
• 50TP2 hits
• 894Signals

Why raw accuracy is the wrong question

A service boasting "85% accurate" tells you nothing if its winners are tiny and its losers are huge. Profitability comes from expectancy — win rate multiplied by average win, minus loss rate multiplied by average loss. A disciplined system at 60.7% win rate and 1.27 average reward-to-risk can compound steadily, while a 90%-accurate one that lets losers run can bleed out.

What a trustworthy accuracy claim looks like

It is public, it includes losers, it states the sample size, and it reports reward-to-risk alongside the hit rate. Anything short of that — screenshots, cherry-picked wins, no sample size — should be treated as marketing, not evidence.

Drishti Pro's numbers, in the open

Every Drishti Pro signal is scored objectively once it closes and stays on the public ledger. The current live figures are a 60.7% win rate at 1.27 average reward-to-risk across 894 signals — a win means the first target (TP1) was hit before the stop, and 50 signals have ridden on to the stretch target TP2. Signals that hit neither stop nor TP1 within 24 hours expire and are excluded from the win rate. You can watch them move in real time on the dashboard, and read the reasoning behind each call.

See also: are crypto signals worth it and how AI crypto signals work.

How the win rate is actually calculated

Drishti Pro's headline number is built from every closed trade, not a curated subset. Since 22 July 2026, the Pro book has produced 894 signals. Of those, 750 closed decisively: price touched either the take-profit or the stop before the 24-hour window ran out. The rest expired untouched and are logged but excluded from the win-rate math.

Among the 750 decisive trades, 455 hit target first and 295 hit stop first. That split produces the 60.7% win rate at 1.27 average reward-to-risk. Because the sample is finite, treat the number as a range, not a fixed point. The 95% confidence interval currently runs 57.1% to 64.1%. A few more closed trades can shift the headline figure inside that band without any change to the underlying strategy.

Long trades versus short trades

A system can look accurate overall while actually being a bet on one market direction. Drishti Pro trades across 11 of its 11-asset book in both directions, and the two sides are tracked separately. Long signals show a 64.3% win rate over 557 trades; short signals show 50.3% over 193. If one side carries most of the edge, that says something about the market regime, not just the model. Check the split for the asset you actually trade before assuming the blended number applies to you. For how this stacks up against published benchmarks elsewhere, see the crypto signal win-rate benchmark.

One book, three asset classes

The 11-asset universe isn't only crypto majors. Alongside BTCUSD, ETHUSD, SOLUSD, XRPUSD, BNBUSD, DOGEUSD, AAVEUSD, and HYPEUSD, the book carries PAXGUSD and XAUTUSD, both tokenized gold, and SLVONUSD, tokenized silver. Metals move on different drivers than crypto majors, so folding them into one blended 60.7% figure hides how accuracy differs by asset class. 11 of the 11 have produced signals so far. Check the per-asset pages before assuming a metals signal behaves like a BTCUSD one.

Inside a signal: features to a published call

Every signal moves through the same pipeline. A tracker runs every 15 minutes, closing any open signal that has hit its target, its stop, or its 24-hour expiry. The generator then builds 144 engineered features per asset from 13 live sources. Those cover derivatives pricing, open interest, order-book depth, funding rates, on-chain flows, and macro data.

• Exchange and derivatives data: Delta Exchange, Binance, Bybit, Deribit, Hyperliquid
• Positioning and flow: CoinGlass, Coinalyze, Farside
• Broader market context: CoinGecko, DefiLlama, Alternative.me, yfinance

Claude, running as Opus 4.8, reads those features at temperature 0 with forced tool-calling, so it returns a structured call, not a hedged paragraph. A code-level risk gate then checks the output: reward-to-risk has to clear 1.0, and the target band is clamped to 1.0-1.5R. That gate bounds the downside on a losing signal relative to its target, and it's documented in full at the methodology page.

Reading a single signal correctly

A published signal states an entry, a stop, and a take-profit, plus the reward-to-risk that produced it. The risk gate has already confirmed that ratio clears the 1.0 floor and sits inside the 1.0-1.5R band before publication. If price reaches target first, the trade counts as a win at that signal's stated R multiple. If the stretch target, TP2, also gets hit, it adds to the 50 count without changing the original win. If price reaches stop first, it's a loss at -1R, full stop. There's no partial credit and no rounding in either direction. That's why the ledger's win rate and average R are numbers the tracker computes automatically, not ones anyone hand-picks.

Where accuracy numbers fall short

The win rate is a lagging measure. It only scores trades that have already closed. A hot or cold streak inside a sample of 894 signals can move 60.7% by several points without the edge itself changing. That's what the 57.1%-64.1% range is for: the interval is the honest statement, the single percentage is the convenient one.

Each signal also carries a self-reported outcome probability, and that confidence is graded separately with a multiclass Brier score once the trade closes. A system can post a fine win rate while still being overconfident or underconfident about its own calls, and the calibration check is what catches that. 50 signals have gone on to hit the stretch target, TP2, tracked apart from the headline win rate.

None of this is a fill. A signal is a call at a price, not an executed order. Slippage, funding costs, and your own position sizing on Delta Exchange India will change your actual result. Past win rate, however it's measured, is not a guarantee of the next signal's outcome. See the full ledger, including every loss, at the track record, or pull the raw data yourself from the public dataset.

See Drishti Pro's live signals Get the app

Educational & illustrative only — not investment advice. Drishti Pro publishes AI-generated trade ideas and their public track record for information. Crypto is volatile and you can lose money. Nothing here is a recommendation to buy or sell any asset. Do your own research.