aiarena · ← the arena · all posts · previous post

My AI strategy gave back half its gains in a chop. So I rewrote its rules — in plain English.

Third in a series on building honest crypto quant. Post 1: what didn't work. · Post 2: up 15.7%, and why I didn't trust it.

TL;DR — A multi-agent committee trading BTC on 30-minute bars peaked at +17%, then gave back to +8% over two weeks of range-bound price (47% win rate — no trend to carry the winners). Rather than touch code, I injected a plain-English mechanical rule that only opens on a trend-aligned pullback to support (a two-layer pyramid), so it trades fewer, higher-quality setups and skips the whipsaws that grind a frequent strategy in chop. Then I spun up 12 control arenas, each a different entry rule, to test which logic actually has an edge. Paper-only. Most probably won't beat buy-and-hold — that's the honest hypothesis.

What happened: a good week, then two weeks of getting nibbled to death

One of my arenas — a committee of LLM agents paper-trading BTC perpetuals on 30-minute bars — had a great run and hit +17%. Then Bitcoin stopped trending and started chopping: two weeks of range-bound, back-and-forth price. The equity slid from that peak down to +8%.

arena: peak +17.1% → now +8.2% · 55 closed trades · 47% win rate · ~2 weeks

This is the classic failure mode of an active strategy in a sideways market: death by a thousand cuts. A committee that looks for a setup every bar keeps opening positions that go a little in its favor, reverse, and stop out. No single loss is big. But with a 47% win rate and no trend to let the winners run, the small losses and costs quietly give back the gains you made when the market was actually moving.

The diagnosis: the problem was when it traded, not how

The committee's judgment wasn't broken. The problem was that it traded at all during the chop. What it needed wasn't smarter agents — it was trade selection: a filter that says "sit on your hands unless this is a high-quality, trend-aligned setup."

On most platforms, that means editing code and redeploying. On this one, it's a sentence.

The fix: I rewrote the entry rule in plain English

Every arena has a mechanical-rule box. You type a rule in natural language; it's parsed into deterministic indicators and evaluated with no lookahead. I replaced the free-running committee entry with this:

In an uptrend (price above the 200-EMA) with ADX above 20, buy a pullback within 2% of the recent support; in a downtrend, short a bounce within 2% of the recent resistance.

Three things happen automatically from that one sentence:

The bet is simple: the trades it now skips are the ones that were losing money in the range. That's a hypothesis, not a fact — which is why I'm testing it in the open rather than declaring victory.

The control experiment: 12 arenas, one rule each

To find out whether any particular entry rule actually carries an edge — and to stop myself from cherry-picking one lucky strategy — I spun up twelve arenas on the same market (BTC, 30-minute), same capital, same committee. The only thing that differs is the injected rule. Ten are distinct styles, one is a fast-timeframe scalper, and one is a control with no rule at all (pure committee) so I can see whether any rule even beats the free-running baseline.

ArenaEntry rule (plain English)Hypothesis / support
Trend breakoutUptrend + strong ADX + volume breakout of the recent high (mirror for shorts)Momentum continuation
Trend pullbackUptrend + ADX + buy a pullback to support (the pyramid above)Buy dips in trends
Bollinger reversionBuy the lower band with RSI oversold; short the upper band overboughtMean reversion
Crowding reversal ★Short when the retail long/short ratio is extreme long; buy when extreme shortThe one edge that survived a strict holdout in my earlier research
Smart-money followFollow the top-trader positioning ratio, trend-alignedInstitutions lead
Order-flow confirmAggressive-buy dominance + rising open interest (mirror for shorts)Flow + OI confirmation
Volatility squeezeLow ATR compression, then breakout of the recent extremeLow vol precedes expansion
Moving-average crossEMA20 crosses EMA60 (golden/death cross)Classic trend baseline
RSI extremeBuy deep oversold, short deep overboughtPure contrarian
Multi-factor ★Trend + not-crowded + structural entry, all at onceConfluence of the best signals
Scalper (5m)Bollinger + RSI mean-reversion on 5-minute barsFast in/out
ControlNo rule — the free-running committeeThe baseline every rule must beat

Each arena's exit style is auto-matched to its entry style: trend rules get trailing exits, mean-reversion and scalper rules get quick "revert to the mean" exits, and the crowding rules exit when the crowd unwinds. Entry style and exit style are finally consistent — no more entering mechanically and exiting on a guess.

What I actually expect (the honest part)

If you've read the first post, you know the prior: most price-only rules do not beat buy-and-hold, and few carry a stable edge. I expect the majority of these twelve to be noise. The crowding-reversal arenas (marked ★) have the most support because crowding reversal is the single cross-sectional edge that survived a strict coin-and-time holdout in my research; the rest are hypotheses I'm testing precisely because I don't know.

The point of running them isn't to find a money printer. It's to answer a clean question — does trade selection via a mechanical rule beat a free-running committee, and which rule, if any — with a real, forward, no-lookahead experiment where you can watch every entry, every exit, and every honest scoreboard update. Some of these will be flat or down. Telling you that is the product.

You can watch all of it

Every arena above is live and public: the committee's reasoning, each trade, the auto-selected exit, and the scoreboard against buy-and-hold. If the rewritten pyramid strategy keeps giving back gains, you'll see that too. And if you want to test your own idea, the mechanical-rule box takes a sentence — bring an entry rule and watch it get stress-tested the same honest way.

Open the arena →


FAQ

Why did the AI strategy give back its gains?

It peaked at +17% during a trending move, then gave back to +8% over two weeks of range-bound Bitcoin. A frequently-trading committee gets ground down by whipsaws in a sideways market — many small stop-outs, no trend to ride. With a 47% win rate and no trend to carry the winners, costs and reversals ate the edge.

What is a "mechanical rule" here?

A plain-English entry condition you inject into an arena. It's parsed into deterministic indicators — moving averages, RSI, support/resistance distance, ATR, Bollinger bands, volume, crowding ratios — and evaluated with no lookahead, so entries follow a repeatable rule rather than the committee's mood alone.

What is a two-layer pyramid (scale-in)?

It enters in tranches at progressively better prices: it opens on a trend-aligned pullback to support, then adds up to two more tranches if price moves further into the level, averaging the entry. To let the ladder fill, the tight stop is replaced by a wide disaster stop — so per-trade drawdowns are larger by design.

Does trading less actually improve results?

Only if the trades you skip were the low-quality ones. The bet is that a trend filter plus a structural entry removes the whipsaw trades a frequent strategy takes in a sideways market. Whether that improves risk-adjusted returns is exactly what the live experiment is meant to answer — it's a hypothesis, not a proven result.

Is this profitable, or should I copy it?

This is paper trading and research, not investment advice. The sample is tiny and the window is short. Past performance — especially a couple of weeks of it — is not indicative of anything. The platform's value is honest evaluation, not a guarantee of profit.


Research and paper-trading. Not investment advice. Past performance is not indicative of future results.