This website uses cookies

Read our Privacy policy and Terms of use for more information.

Welcome back to the Miles Deutscher weekly newsletter! 

Last week I had 2 Claude models build 200 Bitcoin trading strategies. Then I locked away a full year of price data so neither of them could peek at it. 

The strategy that looked best on paper turned $10,000 into $181,057. When I ran it on the hidden year, it made 3%. 

I'll show you exactly how I ran the test, with the prompts so you can try it yourself. I've also got a template I use to get Claude to check its own work before it answers. 

In markets, crypto priced Anthropic's IPO before Wall Street did, while Bitcoin fell 13% and stocks went up. I'll go through the theories for the selling and which ones we can actually check. 

After our main breakdown, stick around for AI Skills & Prompts, this week's market update, and the MHC Recap of setups from inside the room.

AI X TRADING: This week's AI workflow built for traders

I Hid a Year From Claude. Its Best Strategy Flopped.

Every week someone posts an AI trading bot that "made 800%", and almost nobody shows you what it does on data it's never seen.

So I ran that test myself and filmed the whole thing.

I put 2 of Claude's best models head-to-head: Fable 5.2 (the new one) against Opus 5.5. Each had to build 100 Bitcoin strategies on the 4-hour chart.

The setup

Before either model saw any data, Claude set the rules.

→ The data. BTC/USDT 4-hour candles from Bybit. April 2020 to October 2025was the practice period: 12,240 candles.

→ The exam. The next 12 months, 2,191 candles, stayed hidden. Neither model could see a single one.

→ Real costs. Trading fees and slippage were baked in, so the results are close to what you'd actually get.

→ No cheating. 2 separate Claude Code sessions, the same prompt, each locked in its own folder.

The hidden data is what people call "out-of-sample". It's the best way I know to catch overfitting, which is when a strategy basically memorises the past and then falls apart on anything new.

What happened

Fable finished in 13 minutes. Opus took 17.

Fable's top pick only trades breakouts when 2 of 3 slow trend signals agree. On paper: +791%. $10K would have become about $89K. Opus's pick returned +331%.

Then I unlocked the hidden year.

The strategy with the best backtest of all 200 (a 50-day moving average one) made +3.12%.

The winner on the hidden year was an Opus strategy that looked pretty bad on paper. It made +84%.

What surprised me more was that Bitcoin fell 29.44% in that hidden year, and 195of the 200 strategies beat holding it. 148 of them made money.

Before you get excited, though, look at the other side of it.

Over the full 2020 to 2025 paper period, only 2 of the 200 strategies beat just holding Bitcoin. And in a falling year, any strategy that sits out of the market half the time is going to beat holding pretty much by default.

How to read a backtest (3 numbers)

→ Max drawdown. The worst drop from a peak. Holding BTC made 1,600% on paper, but at one point you were down 77%. Fable's pick capped that at 49%. Honestly, most people can't sit through a 77% drawdown, so they never end up seeing that 1,600%.

→ Sharpe ratio. Return divided by how bumpy the ride was. Above 1 is good. Above 2 or 3 usually means luck, a bug, or too few trades.

→ Trade count. 1 Opus strategy scored a 2.42 Sharpe in the hidden year. It took 4trades, which is nowhere near enough to know if it's real.

Do this before next Tuesday (under 30 minutes)

You don't need 200 strategies. 20 is plenty for a first run. Open Claude Code (or Claude with file access) and paste these in order.

Prompt 1: lock the exam

You are my quant research assistant. Download BTC/USDT 4-hour candles from[EXCHANGE] from [START DATE] to today. Verify the data: no gaps, no duplicate candles, timestamps in UTC.


Split it in two:

1. Practice set: everything up to [DATE 12 MONTHS AGO].

2. Exam set: the last 12 months. Move it into a separate folder called /hidden. Do not open, read, or summarise anything in /hidden until I say "unlock".


Build a simple backtester that includes [0.06%] fees per side and [0.05%]slippage, and only uses data up to the current candle (no lookahead). Show me the folder structure and a sample trade log before we continue.

Prompt 2: build the strategies

Using only the practice set, create 20 distinct rule-based strategies for BTC on the 4-hour chart. Spread them across trend following, mean reversion, breakout, momentum, volatility, and regime filters.

Optimise for robustness. A great practice-set number means nothing if it breaks on new data. A strategy that only works with one exact setting is overfit; so test each across nearby settings, every individual year, and double fees.

Write a file with your top 5 ranked. For each: the rules in plain English, return, max drawdown, Sharpe, number of trades, why you expect it to hold up on unseen data, and its biggest risk. Also list the ideas you tried and abandoned, and why.

Prompt 3: the referee (new chat)

Two AI sessions built trading strategies on a practice set. Act as the referee. First, prove nobody cheated: audit the backtester for lookahead bias, check fees and slippage are applied, and confirm /hidden was never opened.

Then "unlock" and run every strategy on the hidden set. Report a table: strategy, practice return, hidden return, max drawdown, Sharpe, trades, and buy-and-holdBTC over the same window.

Flag any Sharpe above 2 with fewer than 30 trades as unreliable. Tell me which strategies held up, which collapsed, and whether any beat buy-and-hold on both sets. If the data is insufficient to judge, say so.

If one survives, get Claude to rewrite it as a TradingView Pine Script so you can check it on TradingView's own data before it goes anywhere near real money.

Where this goes wrong

My test only had 1 hidden year, and it was a falling one, so anything that sat in cash looked great. Run the same test on a rising year, and the leaderboard could look completely different.

Also watch for the "lookahead" bug, where a strategy accidentally uses a candle that hasn't closed yet. It's the most common reason a backtest looks too good, and it's why I give the referee a fresh chat.

Even when something passes, I don't go straight to real money. I run it on a demo account or a small sub-account for a few months and only add capital while it behaves the way the test said it would.

The hard part is knowing what kind of market you're in, and a backtest can't tell you that. Our analysts go back and forth on it every day inside MHC.

AI SKILLS & PROMPTS: Manually curated prompts, skills, workflows & more

The Template That Stops Claude Guessing

A lot of bad AI answers come from 1-line prompts.

If you just ask "is this a good trade?", Claude has almost nothing to go on. So it fills the gaps with confident guesses, and you can't tell which bits are guesses.

We turned Anthropic's prompting guidance for the new Claude models into 1 fill-in-the-blanks template. I shared the full version here.

Below I've filled it in for something most traders could use: a second opinion on a trade plan.

→ The trade plan reviewer

Paste this into Claude or ChatGPT:

<role> You are a risk manager with 15 years of crypto trading experience. Your strength is finding the weak point in a trade plan before money goes in. </role>

<context> I trade [SPOT / PERPS] with a [$ ACCOUNT SIZE] account. I max risk [1%]per trade. My usual mistake: [e.g. moving my stop, sizing up after a loss].</context>

<task> Review the trade plan below and tell me whether it's ready to place. Success means I know exactly what would make me wrong and what I lose if I am.</task>

<inputs>

<data source="my plan"> [TICKER, entry, stop, targets, timeframe, reason] </data><data source=" chart notes"> [key levels, trend, funding, upcoming events] </data>Use only these inputs for facts. If something is missing, ask. </inputs>

<thinking> Find the strongest argument against this trade, then check the maths. Show me only the result plus a one-line why for each point. </thinking>


<criteria> Before you reply, check: position size matches my max risk, the stop is at a level that actually invalidates the idea, reward to risk is at least 2, and no event in the next 48 hours could gap through my stop. </criteria>

<uncertainty> If you're not sure, say so and give a confidence level. Never invent prices, levels, news, or data. </uncertainty>

<output> A verdict (place it / fix it / skip it), position size, the "I'm wrong if" line, and the 1 change that would improve the plan most. </output>

What it forces you to do:

  • Write the plan down before you trade it (a lot of bad trades don't survive this step)

  • Give Claude only your inputs, so it can't fill gaps with made-up levels

  • Make it run its own QA pass against your risk rules before it answers

The catch: it only works if your inputs are honest. If you paste in a plan you'vealready decided to take, you'll just argue with whatever it says. So run it beforeyou're attached to the trade.

Do this: save the filled template as a note, or as a Claude Skill if you use them. Run it on your next 3 trades and keep every verdict. After a week, you'll know whether "fix it" trades are your worst ones.

One more tip if you're on Fable 5.2. Anthropic's own guide says you get most ofits ability at High effort, so there's no need to burn tokens on Max for a trade review. It also sometimes stops early on long tasks, so I add this line: "Finish the whole task before you reply."

MARKETS: The catalysts and moves that matter this week

Market Update

Bitcoin lost 13% last week while stocks went up, and the most talked-about IPO is SpaceX landing this week.

There are plenty of theories going around about the selling, and a few of them we can actually check.

→ Bitcoin overview

BTC closed the week at $65k, down 12.7%. That's the 6th red weekly candle in a row. Over the weekend, it tagged about $62,000, the lowest since July, and $2B of positions were liquidated across crypto in 24 hours.

It's now well below the $75,456 average cost of Strategy, the biggest corporate holder in the market.

My read: the push back into the $80Ks was a bull trap. For me, losing the $74Krange low was what confirmed the bear case. $60K is the major support, and I expect it to hold.

My plan: if the 4-hour chart builds an uptrend, I'll look for a swing back into the high $70Ks. If it doesn't, I'll bid on the weekly close to $60K with my invalidation just below.

If we lose $60K, I'd treat it like losing $40K in 2022, which is capitulation territory. Until then, I'm only laddering into BTC, HYPE, and ZEC and waiting for strength in everything else.

→ Key takeaways: the Anthropic IPO and your bags

1. Crypto priced Anthropic first.

Anthropic prices on Wednesday and starts trading as ANT on Thursday, with a range of $70 to $78. That's up to a $1.4T valuation. Revenue for the first 9 months of the year was $33.2B, up 1,206%.

A pre-IPO perp on Hyperliquid closed at $103.93 on Nov 11, 33% above the top ofthe range.

Why it matters for your bags: crypto is now doing live price discovery on stocks that haven't even listed yet, which is a genuinely useful thing for Hyperliquid. That perp allows 10x leverage, though, so it could get messy if the real price lands far from it.

2. Retail gets an unusually big slice.

About 20% of the shares are going to retail, against a typical 5 to 10%. Even CZ floated IPO cash-raising as one reason for the slide.

Why it matters for your bags: the theory is that people are selling crypto to fund their allocation. It's possible, but it's also a pretty convenient excuse.

3. Last time, the data didn't back the story.

We heard the same theory before SpaceX listed in June, when BTC fell about 16%and briefly broke $60K. But CoinDesk found no abnormal stablecoin cash-out. The clear seller back then was the ETFs.

Why it matters for your bags: we can actually test this one. If IPO cash was the problem, the pressure should fade after Thursday. If the selling carries on into next week, it's probably something else.

4. The ETFs are selling again.

Spot BTC ETFs lost $1.07B last week, the 2nd $1B week in a row. That's 7 straight days of outflows. Meanwhile stablecoin supply had its worst month since 2022, with about $15B redeemed.

Why it matters for your bags: stablecoins are basically the cash crypto buys with, and that pile is shrinking. On top of that, the ETFs were the biggest dip buyers of the year, and now they're selling. I want to see both of those turn before I trust a bounce.

5. Stocks went up while Bitcoin fell.

From Monday to Friday, the S&P 500 rose 2.5% while BTC fell 8.7%. All of the drop happened outside US trading hours.

Why it matters for your bags: last week BTC fell less than the Nasdaq. This week it fell while the Nasdaq rallied, so that relative strength is gone and most of the weakness looks specific to crypto.

6. The one bullish signal: leverage got flushed.

Bitcoin open interest on Binance is back under its 180-day average. We checked every flush like this since 2022. BTC was higher 90 days later every time, by 33%to 76%.

Less leverage means fewer liquidation cascades, and that's usually part of how a bottom forms. It just doesn't tell you when we'llget there.

→ The macro backdrop got worse

October CPI came in hot at +0.4% against 0.3% expected. Markets now give a December 9 rate hike about a 90% chance. The 10-year yield sits at 5.15%, and the dollar is at a 19-month high.

For a proper move up, I think crypto needs the Fed to pivot, and right now it'sheading the other way.

MHC RECAP: Setups and wins from inside Miles High Club

This Week Inside MHC

→ Miles on $BTC ( ▲ 4.93% )

My levels from Sunday. The weekly close had to hold above $62K, and it did.$68K to $70K is the first real reclaim. Until then, every bounce is a lower high. Lose $62K and the July low near $57.7K is next.

The plan from Markets stands: bid the weekly close near $60K, invalidation just below. I'm not looking to be a hero here.

→ Fabian and Paradise on $ZEC ( ▲ 4.55% )

Fabian's buy zone is $750 to $890 (the old consolidation above the prior highs, plus the 200-day SMA and EMA bands), ideally on a big liquidation spike. His other trigger is a convincing reclaim of $1,050 to $1,100.

On Sunday, he started nibbling just below $900. He plans to accumulate from$800 to $1,000 over weeks or months. Losing the $700 to $800 range is hishigher-timeframe invalidation.

Paradise got there on his own, nibbling spot in his $900 to $1,000 zone on a 4Hdouble divergence.

→ surfXBT on $BTC ( ▲ 4.93% )

With his divergences invalidated, Surf treats this as a proper breakdown and a textbook 4H trend continuation. He'll short any bounce into $71K, as long as the4H trend agrees.

So that's 3 different plans on the same chart, and each one has a clear level where it's wrong.

→ Fabian on $KLED ( ▲ 2.86% )

KLED is back to about a $10M market cap, a level Fabian sees as long-term value. He's swung it for 2x+ from here a couple of times. By his NAV maths, this price implies under about $200M of company equity, below Kled's last raise. He'saccumulating slowly, on a multi-month view.

→ Room win of the week

chickennet managed to secure plenty of profits on this recent move through multiple shorts and took profit in stages on the way down.

Remember: not financial advice; always do your own due diligence.

CONTENT PIPELINE: Latest content from Miles Deutscher Finance

New From Miles

YouTube: [I built a Fable 5.2 trading bot, then tested 200 strategies on a year I had never seen]. Watch now →

𝕏: [My thoughts on the current state of crypto and markets, and exactly whatI'm watching next]. Read now →

Skool: New prompts, workflows, research & more. Join here (100% free) →

MHC: Live setups and analysis from our traders, every day. Join MHC →

See you next Tuesday,
Miles Deutscher & team

Want to sponsor or partner with us? Reach out at [email protected]