AI prediction trading: $10,000 becomes a reported $10,150

I would judge Champ's AI trading experiment by the $150 reported gain, not by its winning bets. A $10,000 bankroll became a displayed $10,150 over 7 days, after direct AI picks, copy trading and paired-market arbitrage. That is a modest result from a short test, not evidence of dependable income.

What AI prediction trading did

Claude helped with decisions on Kalshi and Polymarket, where traders buy contracts tied to outcomes such as an election or a match. Prices move before the event finishes, allowing traders to sell before settlement, when the platform determines the result and pays winning contracts.

The initial search was for trades resolving within the next 24 hours. That constraint ruled out opportunities that might take longer and narrowed the choices to events settling soon. A promising trade was not necessarily suitable for a week-long experiment.

Claude said it could not read Kalshi live, so its assessments relied on supplied screenshots. A person selected the information and placed the orders. The model did not independently manage the account.

I keep those jobs separate when assessing an AI trading claim. Interpreting screenshots differs from copying a wallet or executing paired orders. The experiment moved through all of them: directional bets that depended on a chosen outcome, wallet assessments, then software built with Claude to trade price differences. Evidence that a model can help implement a rule does not establish that it can forecast winners.

How contract prices turn into gains and losses

In the example involving Mamdani winning Time's Person of the Year, a Yes contract cost $0.32. It would pay $1 if he won and become worthless if he did not.

That price suggested an approximate market-implied probability of 32%. I would use it as a shorthand for what traders were willing to pay, not an objective measurement of the future. Trading costs and limited liquidity can make the quoted price less useful than it first appears.

Selling before settlement changes the calculation. A contract bought for $0.32 and sold for $0.50 gains $0.18 per contract before costs. That is a price difference per share, not a percentage return on the investment. The purchase price determines the capital committed; the payout determines what a winning contract returns.

High-priced contracts can have an awkward payoff: a small gain if they win and a much larger loss if they fail. A high probability of winning does not, by itself, make the price attractive.

A limit order sets the price a trader will accept. It can prevent an unexpectedly expensive purchase, but another trader must agree to fill it. Before trusting a recommendation, I would look for a reason the contract is mispriced, rather than just a plausible story about the event, and check whether the trade still works after costs.

The first AI picks split the results

Claude's first recommendation was that Kevin Warsh would say the word "oil" during the address. It estimated a 65 to 70% chance, reasoning that energy and inflation could prompt discussion. That was the model's forecast.

For a $10,000 bankroll, Claude suggested a stake between 500 and 750. The chosen stake was $500. A separate AI analyzer recommended betting that somebody would dissent at the Fed meeting. Claude agreed but preferred a cheaper entry: a limit order for $510 worth at $0.60, with a stated payout of $850 if the bet won.

The reported vote was 12-0, so the dissent position lost. The word bet won, paying $784 on the $500 position. The later accounting recorded $284 of profit against a $510 loss on the dissent trade.

I cannot judge the accuracy of a probability estimate from one win. An event can occur after a poor forecast, and a sound forecast can lose. A useful assessment needs a record of estimates and outcomes. These bets also concerned the same broad event, so counting them as separate contracts would overstate their diversification.

The creator claimed that almost 85% of people lose money trading these markets and that bots account for much of the profitable activity. Those were claims, not independently verified statistics. The experiment itself established something narrower: its first AI picks produced a net loss.

Copy trading made exposure grow quickly

Copy trading automatically follows another trader's transactions. Claude approved StudyPrediction and another profile, and $5,000 went into a Telegram copy trading tool.

Within roughly 30 minutes, the tool had copied $1,000 worth of trades. Exposure later reached $2,000, at which point the system was paused while the existing positions played out.

The positions included Bayern winning at $0.94, Chelsea losing at $0.62 and over 4.5 goals in the Bayern match at $0.99. Other bets covered both teams scoring in an AC Monza game and an AS Monaco game ending in a draw. The account owner said they did not follow soccer and found the pace stressful.

I see the problem here as funding automation before deciding what it may do. The allocation did not establish how quickly the tool should deploy money, which markets it should enter or how much related exposure it could accumulate.

Followers may also get a different price or use a different stake from the original wallet. Other positions may offset that wallet's risk without being apparent in a copied transaction. Copying an order does not necessarily copy the strategy behind it.

The expensive contracts left little profit available on each winning share, while losing shares could forfeit their purchase price. Several wins therefore need not cover a few losses. At review, some positions had won, some had lost and one remained open. The creator's assessment was that copying had not produced a useful result for the challenge.

Arbitrage changed the source of the hoped-for return

Prediction-market arbitrage pairs opposite outcomes across platforms. Buying Yes on one and No on another can leave one winning contract at settlement, provided the markets cover the same event under compatible rules. If the combined purchase cost is below the winning payout, the difference is a potential return before costs.

I would check the rules before trusting the word arbitrage. Similar market titles can hide different definitions, deadlines or resolution sources. If an interpretation allows both positions to lose, the pair is not a complete hedge.

An election market supplied the initial explanation, followed by a selected market on Democrats winning the New Hampshire Senate race in 2026. Its displayed spread was 4.17%, described by the creator as about 3% after fees. The planned allocation was $6,000 to make almost 200.

The executed pair was described as Yes on Polymarket and No on Polymarket US, with about $5,000 on the former and about $900 on the latter. Those venues differed from the initial Kalshi and Polymarket comparison. The contracts actually bought determine the hedge.

The creator described the profit as locked in after placing the pair. I read that claim as conditional on both sides executing correctly and the platforms settling as expected. It removes the intended directional exposure, not every possible loss.

The creator also disclosed helping build the scanner that found the opportunities, a financial-interest consideration when evaluating the performance account. The video lacks a complete trade ledger and full fee and settlement-rule breakdown.

The paired-trade workflow and its checks

Before automation, a scanner found pricing differences, software sized and placed the selected pair, and an alert indicated a possible exit. I would understand that process before letting software repeat it unattended.

1. Compare Yes and No contracts for the same event across the available platforms.
2. Sort opportunities by volatility to find spreads that have opened and closed frequently.
3. Select the paired market to inspect.
4. Set the total amount to allocate to the pair.
5. Review the automatically sized positions before submitting the paired order.
6. Monitor the sell notification for the open pair.
7. Review the exit prices before selling both positions.

Here, volatility meant how often the gap opened and closed in the tool's display. The selected opportunity had reportedly done so 5 times in the preceding week. That history informed the search for a trade that could close within the challenge.

The creator explained that converging prices could allow both positions to be sold before the underlying event resolved. Historical convergence does not guarantee another exit, however, and both sides need buyers at usable prices.

I would calculate sale proceeds rather than rely on lines meeting on a chart. Executable prices and selling costs determine the recovery. Selling the wrong quantities can also leave an unintended directional position.

Money committed to an open pair is unavailable for other opportunities. If the gap persists, the choice may be to wait for settlement or accept a different exit price. Liquidity, the ability to trade without substantially moving the price, affects that choice.

Account safety matters before funding any tool. Check the site address and the permissions requested before connecting a wallet, and use a small amount when evaluating unfamiliar software. Trading access creates an operational risk as well as a financial one.

Automation added speed and another failure mode

Matthew helped with the automated strategy and said his own version had been running for about a month and working well. That was his performance account, not a measured result from this experiment.

After hours of prompting, Claude Code produced a bot for an owner who said they had no coding experience. It connected to Polymarket, Polymarket US and Kalshi, scanned opportunities from the existing tool and applied criteria including a size below 4K and expected value above $100.

The interface separated realized profit and loss from "locked EV." Expected value, or EV, estimates a benefit under the strategy's assumptions. An open trade's estimated benefit is not the same accounting category as a completed result.

I can see a use for automating repetitive checks, but errors repeat faster too. Faulty market matching, sizing or balance handling may affect multiple trades before anyone notices. Running code and a polished dashboard do not establish dependable execution.

The bot entered a JD Vance pair described as Kalshi Yes and Polymarket No. The creator claimed at least another $100 of EV, not money already available to withdraw.

A maximum size alone cannot protect a paired strategy. The system needs to detect a failed order on one side, prevent repeated submissions and reconcile its records with the accounts. If only one leg fills, the hedge becomes a directional bet.

I would also want a stop mechanism that disables new orders while retaining a clear view of existing positions. Pausing a program does not close what it has already bought.

Judge the result by the accounting

The displayed balances moved from an early loss to a small reported gain. Their accounting labels were not identical.

The starting and ending figures give a reported difference of $150. I would not substitute a percentage from an individual arbitrage for the bankroll's overall result: other methods lost money, and some capital remained committed.

On the final day, a sell notification arrived for the earlier pair as its price lines converged. The creator later claimed that one arbitrage sold for like 6% in 2 days. That is a stated trade result, separate from the balances in the table.

The creator rejected asking AI for random directional bets as a credible way to get rich overnight, while expressing more optimism about capturing price differences. The reported experience supports the first judgement; the second needs evidence from more trades and changing conditions.

Changing strategies during the week prevents the test from isolating any single method. A short result mixes execution and luck with strategy selection and accounting choices. A positive ending balance does not establish how the bot would handle prolonged price gaps or disputed outcomes.

For another AI trading claim, I would start with completed trades: purchase costs, sale proceeds and fees. Keep open estimates separate, then check settlement terms and whether both sides filled in the intended quantities. The unresolved question is how the bot handles a failed order that leaves only one side of the hedge open.