Can AI Trading Bots Beat Bitcoin After Costs?

Quick decision

Do not choose a bot from one dated leaderboard

The September 20 leader was only 0.44 percentage points above the source BTC reference; unequal timing and asset permissions weaken that comparison.

Separate paper results from real execution

The external September edition and CoinNudge's forward paper accounts simulate fills; neither proves a real exchange profit.

Check turnover before prediction claims

In the separate developer-reported Jev case, fees accounted for 52.4% of the reported one-day loss; this is a self-report, not a reconciled ledger.

AI trading bot profit: which results are live decisions and which are only backtests?

Classify a bot's result as historical rule replay, forward paper decisions or reconciled exchange execution before judging whether it beat Bitcoin. These evidence types cannot be pooled into one AI success rate.

An AI-written rule can be evaluated by ordinary code. That evaluates the rule and fill assumptions, not a model’s autonomous judgment. Likewise, JEV may classify or review a candidate snapshot without producing a verified strategy edge. An LLM confidence score is not a calibrated trade win probability.

The dated external figures remain below, with their original source and comparison limits. The internal evidence is a separate one-year 15-minute BTC experiment, with 35,040 complete intervals derived from 5-minute history. Its primary $10,000 account finished at $9,327.83. Do not put this result beside a 2026 paper model and imply equal periods or fair rankings.

For a practical claim record, write down source, observation cutoff, starting capital, asset permissions, who decided, whether fills were simulated, and what can be downloaded. If a field is unknown, retain unknown. The next section tests costs within the historical replay, not across the external and local accounts.

Verify: Inspect current AI paper accounts · Internal rule replay: coverage, trades and fees

Did the internal AI-strategy test remain profitable after a cost stress?

No: the frozen internal rule lost 6.72% over 819 trades at primary costs and 12.42% when both fees and adverse slippage were doubled. The table's shortened 'Doubled fee' label refers to that full doubled-cost scenario, not a fee-only intervention.

A model can look good by avoiding a falling market while losing cash itself. It can also beat passive BTC with substantially lower time in market. Check absolute final capital, risk, drawdown, trading costs and exposure alongside the benchmark difference. This unfavorable internal example stays public precisely because it tests a claim rather than promotes a model.

Model inference and failure costs must be measured in the actual forward deployment. This old rule replay has no per-candle model bill and cannot be assigned fabricated token expenses. For paper accounts, request a dated decision ledger, actual provider usage, timeouts/retries and rejected or delayed orders before discussing net economic value.

To evaluate a different bot, subtract software and inference expense only where those charges are not already included in reported net equity. This replay embeds slippage in fill prices and lists commissions separately. Missing model bills stay unmeasured, not zero-cost claims.

What should you ask before paying for an AI crypto trading bot?

Ask for the common start and end, account deposits and withdrawals, open losses, closed trades, actual maker/taker fees, funding, slippage and API costs. A screenshot of winning fills omits positions still losing; a high win rate can hide a rare loss that overwhelms many wins. Reconcile account equity and cash, not just a headline percentage.

Check safeguards outside the model: independent maximum position risk, stale-data rejection, order acknowledgment, duplicate-order prevention, stop monitoring during provider outages and revocable trading credentials. A reasoning model is not a substitute for those controls. Paper stop-first handling or reference fills still differ from exchange depth.

Use a frozen rule baseline and a prospective period with unseen decisions; log prompt/model versions and all rejected candidates. If you change a rule after looking at losses, start a new version rather than rewrite history. Review the current paper ledger for transparency, then test a user-specified supported strategy as a separate experiment.

Request one full dated statement and its corresponding order/fill export, reproduce net profit including open marks, then verify a revocation or pause path without granting new permissions. If only a winning screenshot or brand ranking is available, postpone the performance-based purchase; that is an unverified claim, not proof of fraud.

Did any AI trading bot beat Bitcoin in the dated September benchmark?

Yes, against the source's published reference: 2 of 17 paper models were ahead at the September 20, 2026 snapshot. The leader's 5.63% versus displayed BTC 5.20% produced a 0.44-percentage-point excess from unrounded values. Unequal dates, mixed crypto/equity permissions and an unfinished season mean this is not a common-window alpha estimate.

At the September 20 cutoff, Mistral Medium 3.5 led at 5.63% versus the displayed BTC reference of 5.20%, with a 0.44-percentage-point difference from unrounded data. The season was unfinished at that observation and the BTC baseline began a day later; this is a dated comparison, not today's lead.

11 of 17 models were above zero at the September 20 cutoff, but profitability and beating Bitcoin are different tests. A model can make money and still underperform simply holding BTC. A short paper season cannot establish persistence, capacity or real execution quality.

The apparent 0.43-point difference from subtracting the rounded display values is not the computation rule. Use the unrounded return fields in the linked JSON for excess return and strictly-greater counts; rounding can otherwise change a near-tie verdict.

Verify: TradeRank benchmark · Derived snapshot JSON

What does the September 20 AI paper leaderboard show against BTC?

The following 17 rows are the frozen September 20, 2026 external edition, not today's standings or CoinNudge's current accounts. Excess return uses unrounded source returns minus its BTC reference in percentage points; 'Ahead now' and 'current' in preserved table/chart labels mean at that edition's cutoff only.

The source’s reported trade count can be zero while return is non-zero, so we preserve the field as “reported trades” and do not interpret zero as no open position or no activity. Readers should inspect the source ledger before drawing execution conclusions.

Read the first two rows as claims to investigate, not copy instructions. Check unequal benchmark starts and participation dates; drawdown and reported trades are source fields, not a reconstructed complete account path. To see new CoinNudge actions instead, open the separate paper-account tool.

Verify: Open source facts feed · Download derived CSV

What does the evidence say about Jev AI trading crypto?

The available Jev evidence does not establish a fair BTC-beating or consistently profitable record. In the September 20 snapshot Jev joined late, reported 0.00% and zero reported trades; the separate one-day developer experiment used different markets and frequency.

A separate developer reported a 24-hour Jev scalping experiment using $100,000 of simulated capital across three CME micro futures. The author reported 731 trades, a $3,150 loss and $1,650 in fees. By arithmetic, fees represented 52.4% of the reported loss, while the remaining reported pre-fee loss was $1,500. Only 21% of trades were reported profitable after fees.

That experiment is useful because it exposes turnover risk, but it is an author self-report, not a CoinNudge-reconciled ledger. It also mixes Bitcoin, Nasdaq and gold futures, uses 30-second decisions, and is not comparable with TradeRank’s daily investor-style cycle. The correct conclusion is not “Jev fails”; it is “fast decisions did not create an edge in this short setup.”

Treat the two Jev records as separate claim IDs with their own capital, window, market and source reliability. A late-start flat return cannot be scored as a full-season defeat; a 731-trade self-report cannot stand in for a downloadable account audit. Ask for common-window ledger evidence before selecting a model.

Verify: Jev experiment report · CoinNudge Jev evidence guide

How can fees and turnover erase an AI trading bot's apparent edge?

High turnover can consume a small gross edge, but the Jev self-report was negative even before fees. Its reported $1,650 charges across 731 trades average $2.26 per trade; the remaining pre-fee loss was $1,500 on $100,000 paper capital.

TradeRank follows a different regime: one decision cycle every 24 hours and a published 0.1% execution fee. Its displayed returns already reflect its own engine rules, so CoinNudge does not subtract a second hypothetical fee from those results. Comparing a 30-second scalper with a daily portfolio model without adjusting turnover would be misleading.

For any bot, publish gross P&L, explicit fees, spread, slippage, rejected orders and net P&L separately. A high win rate can still lose money when losses are larger, and a low win rate can make money when winners are larger. Net account equity is the governing result.

For screening, write expected gross dollars per completed trade minus entry/exit costs minus financing and allocated inference expense. This is an accounting checklist, not a forecast from the sample. If claimed edge is below a justified cost range, higher trade frequency is not evidence that the loss will be rescued.

Verify: TradeRank methodology · Jev high-frequency self-report

How to compare an AI trading bot with Bitcoin without fooling yourself

A fair benchmark needs the same start and end time, the same starting capital and a clear rule for cash. It should include all costs and value open positions at the same timestamp. If the bot can short, trade equities or rotate into cash while the benchmark is long-only BTC, say so explicitly.

The frozen September 20 source benchmark is informative but not a pure crypto-only contest: its models could trade ten cryptocurrencies and fifty U.S. equities, while BTC was a non-tradeable reference. The Bitcoin baseline began September 13, one day after the season opened. CoinNudge reports that dated comparison with the timing and universe mismatch.

A no-cash-flow comparison starts with R = ending marked equity / beginning equity − 1 for each account at the same two timestamps, then bot R minus benchmark R in percentage points. Deposits require a stated cash-flow-adjusted method; unequal exposure requires a risk comparison rather than calling the difference alpha. No exposure-matched result is calculated for this edition.

Verify: TradeRank competition rules

What can 91 completed model-seasons tell you about AI profitability?

The source's nine completed seasons contain 42 profitable model-seasons out of 91, or 46.2%; that denominator counts model-season outcomes, not independent traders, trades or BTC-beating observations.

It does not answer how many beat Bitcoin, whether the same model won repeatedly, or whether results survive a single common cost and exposure model. The source also discloses archive caveats for Seasons 0, 2 and 7. We preserve those limitations rather than turning the aggregate into a false precision claim.

The important lesson is base-rate discipline: before choosing a model because of one screenshot, compare its result with the roughly 46% historical profitability rate, Bitcoin, cash and a simple deterministic strategy over the same window.

One model may appear in several seasons and models share market conditions, so do not turn 42/91 into a future probability or binomial confidence claim. A buyer needs per-model common-window return and risk records across regimes; this aggregate does not supply them.

Verify: TradeRank public facts and archive caveats

What should an individual investor do with an AI trading leaderboard?

An individual investor should use a dated AI leaderboard to shortlist claims to verify, not copy the highest-ranked model. First separate profit from beating BTC and simulated actions from exchange execution.

If testing your own bot, begin with paper capital, one explicit hypothesis and a benchmark you could actually hold. Freeze the decision rule before the next sample. Record every wait, rejected order, timeout and cost; do not publish only winning trades.

CoinNudge alerts are deterministic market observations and explanations, not autonomous AI execution. They can help a person notice evidence, but they do not inherit the performance of any model in this article.

Before a paper trial, specify what would stop it: missing account reconciliation, risk-limit breaches, unavailable responses or net results failing the chosen benchmark after costs. Choose conditions before observing outcomes rather than retaining only the eventual winning version.

Verify: How to use crypto alerts · Jev integration guide

What research teams should demand before evaluating an AI trader

A research team needs a time-aligned decision, order, fill and account dataset before it can attribute performance to an AI layer. Model and prompt versions alone do not prove execution or profitability.

Run a pre-registered forward protocol with a frozen benchmark. Report model replacements and mid-season joins as new cohorts. Use confidence intervals across days and regimes, not one annualized Sharpe calculated from a short run. Investigate whether returns come from market beta, a persistent short bias, concentration or genuine selection skill.

A useful institutional output is a point-in-time decision dataset: what the model actually knew, what it decided, what was executable and what happened later. Retrospective summaries should not be inserted into the original feature set.

Retain all candidate versions, failed runs and rejected actions. Distinguish event time, when inputs became available and database ingestion; future outcome labels are evaluation, not decision features. Purchase only a fixed edition whose fields and actual coverage meet that task.

Verify: Point-in-time crypto data · CoinNudge data products

CoinNudge method, reproducibility and update policy

We retrieved the CC BY 4.0 public facts feed at 2026-09-20T07:17:51.142202Z. The source snapshot reports its own observation time as 2026-09-20T01:00:04.336Z. We preserve each published return, reported trade count, Sharpe and maximum drawdown, then calculate only excess return, profitability counts and strict BTC-beating counts.

The JSON download publishes formulas, the source hash, non-comparability warning and all derived rows. The CSV contains the September 20 standings. The chart and table come from that same frozen dataset. We do not infer missing trades, reconstruct undisclosed fills or merge the two Jev experiments as though they shared a protocol.

This stable URL is intended for periodic editions. A future refresh must retain the previous evidence directory or issue a versioned replacement, update the source-observation time, and rewrite the direct answer from the new data. It must not silently label an old snapshot “live.”

To reproduce the external comparison, download derived JSON and standings CSV, check the recorded source hash, count unrounded returns strictly above zero and BTC, and compare excess fields with return minus BTC. The internal replay has a separate 819-trade verifier. Passing arithmetic verifies neither undisclosed source fills nor a current refresh.

Verify: Derived JSON · Derived CSV · TradeRank CC BY dataset

Frequently asked questions

Sources and verification

• TradeRank public factsSeptember 20, 2026 source edition, BTC reference and completed-season aggregate; CC BY 4.0.
• TradeRank benchmarkPublished scope, active-season status and data license.
• TradeRank methodologyDecision cycle, asset universe, execution fee and paper-trading rules.
• Obside Jev experimentDeveloper self-report; not independently reconciled by CoinNudge.
• CoinNudge Jev guideModel interface and integration evidence boundaries.

Check the dated evidence without confusing a paper leaderboard with a trading guarantee.

Inspect alerts and research data under their actual coverage before building your own test.