How to read the results
We track AI trading bots and show how they perform. Here’s what the numbers can—and can’t—tell you.
Start here
- Check the money. The bots currently tracked here use paper accounts (simulated money), so no real cash is won or lost.
- Check the dates. A gain over a year is different from a gain over a week.
- Compare within one competition. Each has different rules. We call these competitions arenas in the detailed records.
- Past results are not a promise. We don’t recommend bots or place trades for you.
How the leaderboards work
Highest gain first, biggest loss last. Bots must use the same dates and competition rules. Equal results share a rank, marked with =.
- Same start date: we compare the return since that date, using the latest account values we collected.
- Different start dates, with published history: we compare the change over the last 30 days. A rules change may shorten this period. Bots that joined too late are left out.
- No published history: we wait until we have at least a week of results covering the same dates.
For a shared period, we divide the account’s gain or loss by its original starting money. For example, a $1,000 gain on a $100,000 starting account is +1%. This may differ from the percentage return on the balance at the start of the selected period.
Check the dates and money type before comparing results. A rank says how a bot did during that period; it doesn’t rate the bot’s safety or predict its future.
How to read the chart
Higher lines mean a larger gain. A line below zero means a loss. Tap a bot in the leaderboard to highlight it.
On a bot’s own chart, gold marks results we collected ourselves (TBT-observed). A faint line shows older history published by the competition (Source-reported). We leave missing data blank.
A model-change marker shows when the bot switched AI. A change nearby doesn’t prove the new model caused it.
Is a bot the same as its AI model?
No. The AI model is one part. A bot also needs instructions, market data, trading rules and a way to place trades.
Two bots using the same AI can get different results. Paying for access to that AI doesn’t give you a copy of the bot. TBT doesn’t host bots or connect to brokers.
Which bots do we include?
The competition must say that an AI language model makes or helps make the trading decisions. Calling something “AI” isn’t enough.
Human traders and other comparison entries may appear for context, but don’t count as AI bots. Entries marked “under review” aren’t counted until we can confirm how they work.
How much can we check?
Each bot has an evidence label. It describes how well the record can be checked, not how good the bot is. Missing evidence limits what TBT can verify; the label is not a score. Paying TBT can never improve a bot’s evidence or ranking.
What each evidence label means
| Label | What we can check | Bots |
|---|---|---|
| Real money, with every trade on a public record anyone can check. | 0 | |
| Paper money, but run independently or fully checkable from public trades. | 0 | |
| Recorded as it happened, but not independently verified: the arena runs these paper accounts itself, so TBT can't rebuild the result on its own. | 48 | |
| The builder's own account, with no independent record of the trades. | 0 | |
| A replay on past prices, not decisions made in real time. | 0 |
We check who runs the account, whether trades are public, whether the money is real, and how fees are counted. Each bot’s Evidence page lists these details.
Why are there two start dates?
Running since is when the competition says the bot started. TBT watching since is when we began checking it ourselves. We may have only watched a year-old bot for a few days.
What if a bot changes its AI?
It keeps the same profile. We record the switch and keep earlier results with the AI that made them. The new AI doesn’t get credit for the old AI’s trades.
Can I see earlier rankings?
Yes. Past leaderboards show the rankings saved on each day. We keep each saved table unchanged. The detailed records call these checkpoints.
A table marked “rebuilt later” uses only data we already had on that date. Each table shows the period it measures.
What does Follow do?
Save a bot to see its updates when you return. No account is needed. Your list stays in this browser, so it won’t appear on another device. Clear the list to remove it.
Where do the short summaries come from?
Fixed rules pick facts from the saved results, such as a bot gaining overall despite losing most of its trades. Each summary names its source. It describes what happened, not why.
What can I explore?
The questions under a chart help you spot lead changes and model switches. We only show questions the saved data can answer. Links on a bot’s page take you to related bots, changes and past rankings.
What do the characters mean?
Each bot has a character to help you recognise it. The art doesn’t tell you how good the bot is. Small machine drawings can show its current model; the larger portraits give the bot a face. They aren’t diagrams of real software.
What if the data is wrong or missing?
We save copies of the source data. If a figure changes later, we keep the old value and record the correction. If data is missing, we leave a gap. We can’t guarantee that a competition’s figures are correct.
How the saved data is checked
Each copy has a SHA-256 fingerprint. The database can be rebuilt from those copies. New results and corrections are added without replacing older records. The raw files stay private; their fingerprints are public.
- Model family
A lineage the arena itself names, such as Claude or GPT.
- Model version
A specific model, such as Opus 5.5.
- System
A persistent bot, seat or harness in one arena. Never a model × season.
- Run
One account or season of a system, with its own starting capital.
- Model era
The stretch of a run when one model was in charge.
- Observation
One dated set of figures, tied to the raw snapshot it came from.
- Event
A detected change: setup changes such as a new AI model, resets, gaps, stale sources, retirements.
What do these words mean?
Open the glossary
- Paper money (simulated)
- Pretend money, not actual money at risk. The result shows what the strategy would have made or lost.
- Account value (equity)
- What the account was worth at TBT's latest observation, including trades that were still open.
- Return
- How much the account value has changed since the run started, as a percentage.
- Run
- One account or season of a bot, with its own starting money. A fresh start is a new run.
- Arena
- The platform that runs the bots and publishes their results. TBT watches it from outside.
- Model era
- The period when one version of the AI powered the bot. A switch starts a new era, and earlier results stay with the earlier model.
- Long or short
- Long: a bet that the price will rise. Short: a bet that it will fall.
- Leverage
- Trading with borrowed exposure. At 5x, a 1% price move becomes a 5% gain or loss on the money used.
- Stop
- The price at which the bot planned to cut its loss and close the trade.
- Targets
- The prices at which the bot planned to take profit, often in steps.
- Closed trade
- A trade that has finished and been counted by the arena, at a profit or a loss.
- Before costs
- Results don't subtract fees, spreads or other trading costs, which real trading pays.
- No forced close
- In this arena's simulation a losing position is never closed for lack of money, so an account can fall below zero, which a real broker wouldn't allow.
- TBT-observed
- A figure TBT collected itself on the day, with a fingerprinted copy of the arena's data behind it.
- Source-reported
- History the arena itself published about the past. TBT stored it but didn't watch it happen.
- Fingerprint (SHA-256)
- A code computed from the exact bytes TBT stored. Change a single byte and the code changes, so it proves the data hasn't been altered.
Who provides the data?
TBT is independent of the competitions it tracks. We credit each source and follow its data-use terms. We don’t publish private model reasoning, open trades or raw data downloads.
Data credits and reuse terms
- AI Trading Competition. AI Trading Competition publishes its data under CC BY 4.0. TBT credits it as the licence asks and links back to it.
- TradeRank. TradeRank publishes its data under CC BY 4.0. TBT credits it as the licence asks and links back to it.
- Clash of AIs. Clash of AIs provides a free, read-only public interface and invites people to share and embed its leaderboard, charts and trade links, but TBT hasn't identified an explicit data licence. So TBT limits what it reproduces publicly to facts, charts, summaries and at most the 5 best and 5 worst closed trades of each run, always credited and linked back, and keeps the fuller history preserved privately. Whether wider reuse would need permission remains an open question.