BotVersusBot Explained: How Two AI Trading Bots Compete for Real Money

What BotVersusBot actually is

BotVersusBot is a live documentation project. Two teams — Krypto and Key — each run algorithmic trading strategies through a shared platform called CoinClaw, which sits on top of Binance’s API. The bots trade independently, with no visibility into each other’s positions or code. Both sides post their results publicly.

The competition started in February 2026 with $10,000 each in starting capital (paper trading). Krypto later deployed real money on Binance. Key kept their strategies on paper while rebuilding the approach from scratch.

Everything gets documented: the wins, the bugs, the infrastructure outages, the strategy decisions that looked smart at the time and later didn’t. If a bot loses money because a test corrupted the live cache — that’s an article. If a strategy fails Gate 1 validation after four months of live trading — that’s an article too. The point is to show what algorithmic trading actually looks like, not a highlight reel.

Most trading bot content online skips the parts where things go wrong. We write those parts first.

The two competitors

Krypto: grid trading and DCA accumulation

Krypto’s current bot is V3.7 — a narrow scalper running on BTC/USDC on Binance. The strategy has two components working in parallel:

The grid scalp component places buy and sell orders at fixed price intervals. When BTC moves down within the grid, the bot buys. When it moves back up, it sells. Profit comes from the oscillation. It does not require a directional view on where BTC is heading — just that prices will move back and forth within the grid boundaries. Through June 2026, this component has produced $2.43 in realized gains.

The DCA accumulation component is a different philosophy: when the market drops, buy more. Each time BTC hits a lower level, a new position opens. The assumption is that BTC will eventually recover, and the bot will exit those positions at a profit when it does. The problem with DCA in a prolonged bear market is that the positions keep accumulating and none of them close — because the recovery condition never fires. As of June 22, 2026, Krypto has 8 open DCA positions with -$19.75 in unrealized mark-to-market loss. Realized DCA losses: -$43.98.

Krypto V3.7 overall: $1,157.00 from an April 2026 starting point of $1,682.40 — down 31.2% over 11 weeks.

Key: three bots with different entry conditions

Key rebuilt from scratch in early 2026 and now runs three paper bots simultaneously under the V3.8 label, each with $10,000 in starting capital:

KSQ-V2 (Squeeze Breakout) watches for Bollinger Band compression below a threshold, then enters on breakout. Compression means the market is coiling — often a precursor to a sharp move in either direction. The bot only enters when the compression is clear enough to signal a real breakout, not noise. Since April 2026, that condition has not fired cleanly. KSQ-V2 is flat at $10,000.00.

KBTR-V7 (Trend Rotation) reads the weekly EMA crossover — a 6-week vs 20-week moving average — and rotates between BTC and USDC depending on which is above the other. When the short EMA is below the long EMA, the bot sits in USDC. It has been in USDC since the bear trend set in. Flat at $10,000.00. That is correct behavior: in a confirmed downtrend, not being in the trade is the right call.

KFGC-V8.3 (Fear and Greed Contrarian) enters when the Crypto Fear and Greed Index drops to extreme fear (a reading of 15 or below), subject to a MA200 filter and a fixed position sizing rule (sam=0.5). The index hit extreme fear during the spring 2026 downturn. KFGC-V8.3 entered. As of June 22, it is the only bot in the entire competition that has made money since April — sitting at $10,064.86, up 0.65%.

Key V3.8 combined: $30,064.86 from $30,000.00 — up 0.22%.

How to read the scoreboard

The scoreboard shows each bot’s opening capital, current value, percentage P&L, and status. A few things to understand about those numbers:

Krypto’s figures are live Binance money. When Krypto V3.7 shows -31.2%, that is a real financial loss on real capital. The bot made trades using actual USDC on a live Binance account. The P&L includes exchange fees. Nothing is simulated on Krypto’s side.

Key’s figures are paper trading. The $10,064.86 that KFGC-V8.3 shows is simulated — no real capital was at risk. The bot executed against live market data with the same logic it would use with real money, but if BTC had moved against the position, no real money would have been lost. The trade history and performance metrics are real; the financial consequence is not.

Starting capital varies. Krypto’s April starting point is $1,682.40 because that was the account value at the start of the current competition window. Key’s three bots each started from $10,000.00 in paper capital. You cannot compare percentage P&L directly between bots without accounting for the difference in capital base and the paper vs live distinction.

Unrealized vs realized P&L. Krypto currently shows two separate figures: realized P&L (from closed trades) and unrealized mark-to-market (the current value of open positions). The -$41.55 realized loss is locked in. The -$19.75 unrealized loss on the 8 open DCA positions could improve if BTC recovers, or deepen if it falls further.

What “status” means. A bot marked 🔴 Live is trading in real-time on Binance. A bot marked ⏸️ Waiting is paper trading but in a hold state — conditions for entry have not been met. A bot marked 🟢 Active is currently in an open position.

Paper trading vs live trading: why both matter

Paper trading uses live market data but no real money. The bot runs its actual logic — the same order sizing, the same entry and exit conditions, the same risk management — but instead of sending real orders to an exchange, it simulates what would have happened.

The advantage is that you can test a strategy under real conditions without risking capital. The problem is that paper trading removes the two biggest variables that make live trading hard: execution slippage and the psychological reality of watching real money move.

Slippage is the difference between the price you expected to trade at and the price you actually got. On paper, the bot fills at exactly the target price every time. In live trading, the market has moved slightly by the time your order executes, especially in fast-moving conditions. Over hundreds of trades, that slippage compounds.

For an algorithmic bot, psychology is less relevant — the bot does not panic-sell. But it does matter for the humans watching the bot. Krypto watching a -31.2% drawdown with real money is a different experience than Key watching a paper portfolio. One of the things this competition documents is that difference: whether live exposure changes how a trading system is managed.

This is why BotVersusBot tracks both. Paper P&L tells you whether a strategy has logic worth testing. Live P&L tells you whether it survives the actual market.

The three-gate validation framework

Before a bot goes live with real capital, it goes through three statistical tests. The framework was built after the early bots in the competition — V3.5 and V3.6 — were deployed without any validation and later failed when tested retroactively.

Gate 1: statistical edge. The question is whether the strategy’s performance is distinguishable from random trading. The test is a Monte Carlo permutation: take the bot’s actual trades, randomly shuffle the returns thousands of times, and see where the real performance lands in the distribution. If the real performance is better than 95% of the random shuffles, the strategy has a measurable edge. The threshold is p < 0.05.

V3.5 Grid came back at p=0.938 — meaning 93.8% of random strategies would have done as well or better. That is not a strategy. V3.8 ETH Grid came back at p=0.003. Gate 1 passed.

Gate 2: walk-forward robustness. Backtesting on historical data is easy to game, intentionally or not. You optimize parameters on the data you have, and the result looks good because the strategy was tuned to that specific history. Gate 2 tests whether the strategy works on data it has never seen. V3.8’s walk-forward efficiency ratio was 2.559 — out-of-sample performance was 2.5 times better than in-sample. That is the opposite of overfitting.

Gate 3: regime fit. A strategy that works overall but only in one market condition is not robust. Gate 3 runs the strategy against different market regimes — bull, bear, sideways — and checks whether performance holds across them. V3.8’s bull-regime Sharpe ratio was +0.218. It passed.

Only one bot in the history of this competition has passed all three gates before going live: V3.8 ETH Grid. V3.5, V3.6, and V3.7 all trade or traded real money without completing the full framework. The results reflect that.

For the full methodology, see The Three Gates.

What four months of live data shows

The April through June 2026 window has been a sustained bear market for BTC. It has been the hardest possible test for Krypto’s DCA strategy and the best possible test for Key’s defensive and contrarian approach.

Krypto’s V3.7 kept doing what it was designed to do: accumulate positions on the way down. Eight of those positions are still open. The grid scalp component held up — it has been profitable in aggregate. The DCA component is underwater because the recovery it was waiting for has not arrived.

Key’s three bots responded to the same conditions in three different ways. KBTR-V7 went into USDC and stayed there — correct. KSQ-V2 found no clean setups and did not enter — also correct. KFGC-V8.3 waited for maximum fear and entered when the Fear and Greed Index hit its threshold — and is currently the only profit on the board since April.

This is not proof that Key’s approach is better. It is evidence that Key’s approach suited a specific market regime: prolonged bear, brief extreme-fear spike, no clean breakout conditions. A different four months — say, a strong trending bull market — would probably have produced a very different scoreboard.

Which is the point. No strategy beats every regime. The competition documents what happens when each strategy meets the market it was designed for, and the markets it was not.

Where things stand now and how to follow along

Krypto V3.7 is live on Binance. The 8 open DCA positions are waiting on BTC recovery. Every week that passes without that recovery adds to the unrealized loss. The question for Krypto’s strategy is whether this is a mid-cycle dip or a sustained downtrend that forces those positions to close at a loss.

Key’s V3.8 suite is mostly in wait mode. KFGC-V8.3 is the live trade — currently positive, exit conditions not yet triggered. KSQ-V2 is watching for a squeeze setup. KBTR-V7 is watching the weekly EMAs for a regime change.

The competition has no fixed end date. It runs until the market or the participants produce a definitive result.

To follow along:

• Live scoreboard — current P&L, open positions, bot status. Updated as results come in.
• Weekly recaps — every week (or when something significant happens), we publish a performance summary. The June 2026 catchup covers the full April through June period.
• Strategy explainers — deeper dives on how specific strategies work and why they were built that way. Start with How Grid Trading Works if Krypto’s approach is new to you.
• Engineering posts — when something breaks or gets fixed, that gets documented too. The cache corruption incident from earlier in the competition is a good example of what that looks like.

The scoreboard updates as results come in. Weekly recaps post on Mondays when possible, mid-week when something happens that cannot wait.