Updated daily · as of Oct 28, 2026 · every number sourced

AI trading bot statistics (2026)

Every statistic we can back up about AI trading bots and agents: how many make money, whether they beat Bitcoin, how AI models do with real money, what users pay, and how often advertised returns hold up. Each links to its source and the date it was measured.

Most numbers are computed by us from public chain and venue data; the rest say where they come from. For the full rankings see the verified leaderboard.

+9.1%trading ROI of 321 AI agents on prediction markets (90 days)

31 of 38Alpha Arena model accounts lost money

6 of 8AI apps on Hyperliquid whose users lost money (30 days)

17 of 18advertised returns nobody else can check

Performance: do AI trading bots make money?

Computed by us from public chain and venue data (grade A). These are the only bots whose results anyone can check.

90%
of AI agents with enough activity on prediction markets made money over the last 90 days: 141 of 157 (at least 20 settled bets, $25 settled and active in the last 30 days). Across all 321 active agents, trading ROI was +9.1% before AI-tool costs.AI prediction-market agents · Olas registry · as of Oct 28, 2026
3,875
AI trader agents have registered on-chain in the Olas service registry to bet on Omen and Polymarket. Every bet and payout is on-chain, which is why we can rank them.AI prediction-market agents · as of Oct 28, 2026
2 of 6
Hyperliquid vaults whose leaders say they're AI-run beat BTC over the same dates. 1 lost money, and 1 fell more than 30% from a peak.Leaderboard: AI-labelled vaults · as of Oct 28, 2026
6 of 8
AI trading apps on Hyperliquid had users with a net realized loss over 30 days. Together their 1,561 users traded $24.2M, and 42% of users who closed a trade were in profit.AI trading apps on Hyperliquid · Hyperliquid builder fills · Sep 28, 2026 to Oct 27, 2026

AI models trading and forecasting

Large language models given real money or real forecasting questions, scored on outcomes nobody could know in advance.

4 of 6
AI models lost money in Alpha Arena Season 1, trading $10,000 each of crypto perps on Hyperliquid for 17 days. Best: Qwen3 Max +20.3%; worst: GPT-5 −63.4%.Alpha Arena recomputed · Oct 17 to Nov 3, 2025
5 of 32
Season 1.5 accounts finished with a profit (US stock perps and crypto). Grok 4.20 (Mystery Model) was the only model up across its four accounts, +13.0%.Alpha Arena recomputed · Nov 19 to Dec 3, 2025
11.3%
of the starting money went to trading fees across both seasons: $43k on $380k in 38 accounts, 3.7 times the combined profit of the 7 accounts that finished up.Alpha Arena recomputed
67.5 vs 61.5
Superforecasters still beat the best AI model on ForecastBench, which scores forecasts on questions that resolve in the future (Brier Index, higher is better; best model: o3-2025-04-16-scratchpad).Leaderboard: AI models · ForecastBench · as of Oct 28, 2026

Advertised results: can you trust them?

What AI trading products say about their own returns, against what anyone can check.

17 of 18
advertised AI trading returns we checked can't be verified by anyone outside the company: 11 are claims with no public record and 6 are backtests or simulations. 1 we measured ourselves.Claim check
11
robot signals on real MQL5 accounts with 90%+ algorithmic trading are shown as the venue reports them (grade B: venue-verified, not computed by us, so not ranked).Leaderboard: robot signals · MQL5 Signals

Usage: how much people use trading bots

Fees and volume measure demand. Most bots publish no trading record, so this is the most we can say about them.

$6.4M
in fees users paid 16 Telegram trading bots in 30 days; 72% went to one bot (Terminal). Usage shows demand, not whether users made money.Bot usage stats · DeFiLlama · as of Oct 28, 2026
$62.2M
in fees users paid 25 trading apps and terminals in 30 days, led by Axiom ($29.4M).Bot usage stats · DeFiLlama · as of Oct 28, 2026
25
AI agent protocols on DeFiLlama report fees or deposits; together $937k in fees in 30 days and $30.3M deposited.Bot usage stats · DeFiLlama

Security: what could go wrong

Results don't show what a bot can do with your money. These do.

5 of 41
open-source trading and agent repos we list have published security advisories (usually with a fix: check you run the patched version); 0 are archived or haven't had a commit in 6 months.Open-source repo health · GitHub · OpenSSF Scorecard · as of Oct 28, 2026
6 of 12
official broker AI connectors (MCP servers) let an AI agent place real orders without asking you first.How broker MCPs gate orders
3 of 36
Agent Risk Scores are below 40 out of 100, the band with the most money at risk if an agent is wrong, hacked or prompt-injected.Agent Risk Scores
16
security incidents on trading bots, exchanges, wallets and agent tools are on our tracker, each from an official disclosure.Incident tracker · Telegram bot hacks

How these numbers are made

Evidence grades

A, verified: computed by us from public chain or venue data. The only grade we rank. B, venue-reported: a platform's own numbers on real accounts. C, claimed: a company's statement with no public record. D, backtest: a simulation.

How we verify results · How we score risk

What these numbers don't say

Past results don't predict future ones, and short windows can flatter luck. A bot that made money for others can still lose yours, and usage or GitHub stars measure attention, not skill or safety. Not financial advice.

Cite these statistics

Free to use under CC BY 4.0. Please credit and link: TradingAgentStack, "AI trading bot statistics (2026)", https://tradingagentstack.com/ai-trading-bot-statistics/, accessed Oct 28, 2026.

Raw data: AI agents (JSON) · AI vaults (JSON) · AI apps (JSON) · Alpha Arena (JSON) · bot usage (JSON) · repo health (JSON). Agents can query everything through our MCP server. Run an AI trading agent with a public account? Get it verified.

FAQ

What percentage of AI trading bots make money?

There's no overall number anyone can verify, because most bots publish no trading record. Where we can check: 90% of active AI agents on prediction markets with enough activity made money over 90 days (141 of 157), but users of 6 of 8 AI trading apps on Hyperliquid lost money over 30 days, and 31 of 38 Alpha Arena model accounts lost money.

Do AI trading bots beat Bitcoin?

Rarely, in the data we can check: 2 of 6 Hyperliquid vaults whose leaders say they're AI-run beat BTC over the same dates. Beating BTC matters because holding it costs nothing and needs no bot.

Which AI model is best at trading?

In Alpha Arena, the only public real-money contest between models, Qwen3 Max led Season 1 (+20.3%) and Grok 4.20 (Mystery Model) led Season 1.5 (+13.0%). Seasons lasted two to three weeks, too short to separate skill from luck, and the same model can finish first in one season and near the bottom in the next.

Are AI trading bot win rates real?

Most can't be checked: 17 of the 18 advertised AI trading returns we checked have no public record or come from backtests. A win rate also says nothing without the size of the losses: a bot can win 90% of trades and still lose money.

How much do people spend on Telegram trading bots?

Users paid 16 Telegram trading bots $6.4M in fees over the last 30 days, according to DeFiLlama. Most charge about 1% per trade, so fees track trading volume.

How often are these statistics updated?

Performance, usage and repo health are recomputed every day from public data. Claims and MQL5 signals are checked by hand each week. Every statistic shows its source and date.