Known for his bot-setup tutorials, Miles says the most common question is no longer how to build a bot but which strategy to run, so he frames the episode as a practical hunt for proven approaches and a path to turn the winner into a live system.
The spark comes from a post by algo veteran David Tech listing four free places to borrow strategies; Miles notes he already used one and four, discovered two and three through the post, and argues there is enough free gold to start before paying for curated libraries.
The first pool is TradingView. He browses trending and editor picks on an asset like Bitcoin, opens community scripts and copies their source. Examples include a volume-weighted order-book zone, AMD PO3 and a dual-phase reversal script; some are pure visuals, others already emit entries, stops and signals.
Instead of trading raw indicators, he asks Claude with Fable to rewrite three Pine scripts into testable strategies with risk and trade parameters, keeping them faithful to the original idea while flagging this as a rough conversion that will need deeper work and forward checks later.
The second pool is Stonehill Forex and its NNFX indicator library. He feeds the site to Claude, asks it to decode the method, write a backtest-ready spec per indicator and create one document per strategy; the job runs long because the library is large.
The third pool is QuantConnect. He highlights community league entries such as an out-of-sample 127% in three months, notes the code ships with a free account, then has Claude bulk-download ten strategies into a single file and asks for an accuracy check against the originals.
The fourth pool is Quantpedia. He describes it as an academic curation that distills papers into plain rules, performance and risk notes and downloadable code, browses unlocked entries with strong track records and repeats the same bulk-download trick.
To round out the universe he asks Claude to invent five proven-style strategies from its own reasoning and reach 50 in total. Token usage spikes and hits a temporary limit, so he advises tighter curation at scale and promises the full list of winners plus a build guide inside his free School community so viewers can replicate without burning the same budget.
With a folder of top strategies ready, he opens a fresh chat, connects the folder and gives a detailed prompt: locate the strategy in each file, keep single-asset rules as is, synthesize a rule from author notes when only an indicator exists, handle the NNFX branch, write one spec per strategy, generate engine files and run everything to report the best.
The dashboard shows 325 runs across 50 strategies. The top H4 Binance run prints 3,447% but is flagged for a 64% max drawdown and a low win rate; other high flyers near 2,400%, 2,300% and 1,700% raise similar questions. Claude highlights two Bitcoin daily strategies with profit factors 1.78 and 1.54, high win rates, low drawdowns and a steadier equity curve; SPY prints 621% on one run, yet many entries lose money, proving a neat indicator does not guarantee profit.
Claude's top five picks make the nuance clear: the outlier that would turn 10k into 354k, a second that would turn 10k into 114k and only recently pulled ahead of buy-and-hold, and three more that trail buy-and-hold in nominal terms but lead on risk-adjusted ground. The commentary stresses that a lower nominal winner can be the better portfolio choice when it controls wiggles and drawdowns.
Hands-on TradingView checks then test credibility. One 328% six-year run rests on just four trades, another 263% run with 16.77% drawdown rests on two long momentum trades, both too thin to extrapolate. More active candidates look different: one with 123 wins in 296 trades, 7.43% drawdown and a small edge over buy-and-hold, and another with 45% return, 55% win rate, 8.86% drawdown and a 1.9 profit factor. The latter, a Stonehill-origin strategy that stays active through July, August and September and stays above buy-and-hold, becomes the forward-test candidate.
The close shifts from picking to proving. The suggested path is paper forward testing, then a small live test, then gradual scaling. For execution he sketches an LLM brain connected via MCP to an exchange such as Bybit, either orchestrated directly or via a reliability bridge, plus a planned guide. He notes tweaks like drawdown caps, probing why a strategy works with Claude, and fitting the strategy to a regime view: bull, bear or chop each favors different types, and every bot needs a defined portfolio role before allocating capital. SPY and Ethereum outliers near 2,333% are flagged as further candidates worth similar treatment.
AI commentary
"What struck me in this test was not the raw return column but how that return was made; a 3,447% figure looks thrilling until you read it alongside a 64% drawdown and just four trades in six years."
AI assessment
Steelmanned, the core counterargument is that a leaderboard built on backtests almost guarantees the best backtest wins. A 3,447% outlier, paired with a 64% drawdown on a single venue, timeframe and asset, reads more like overfitting and survivorship framing than proof of edge, and even an out-of-sample window remains a narrow slice of history.
What was not tested matters as much as what was. Two or four trades in six years cannot support a 263% or 328% claim with confidence; spread, slippage, fees, funding and order-book depth are only approximated in a simulator and regimes change. Denser samples such as 123 wins in 296 trades with a 7.43% drawdown give me more to work with than a thin, high-return tail.
Incentives and verifiability sit in the third lens. The episode credits David Tech and funnels viewers toward a free School community, which is transparent but still a funnel; the right response is targeted re-checks. On my list are the exact out-of-sample definition, how profit factor and max drawdown are computed, under what cost assumptions that showcase 127% on QuantConnect was produced, and over which dates and adjustments the buy-and-hold comparison is drawn.
Practically, I would split the audience in two. For an experienced algo tinkerer who will paper-trade for weeks, interrogate why the edge exists with Claude, and test drawdown caps and sizing variants before sizing up, this collection is fertile ground. For someone expecting plug-and-play income who reads a four-trade curve as a promise, the same collection is an expensive lesson; the first question should be which regime you believe the strategy fits and what portfolio role it is meant to play.
Sources
9 links; 1 of them also cited by 1 other story. Stories sharing a link do not confirm each other; a source's origin is not inferred from how often it is cited.
- @youtube.com YouTube — 50 Bots Tested Episode (Miles Deutscher Finance)
- @tradingview.com https://www.tradingview.com/script/ubNhdO2q-Grand-Master-s-Candlestick-Dominance-ATR-Enhanced/
- @stonehillforex.com https://stonehillforex.com/indicator-library/
- @quantpedia.com https://quantpedia.com/how-it-works/
- @quantconnect.com https://www.quantconnect.com/league/2025-q2/
- @quantparadox.com https://www.quantparadox.com/blog/mt5-strategy-tester-limitations
- @alfatactix.com https://alfatactix.com/academy/mql5-ea/strategy-tester-vs-live-mt5
- @monkeytrade.com https://www.monkeytrade.com/learn/trading-psychology/survivorship-bias-in-trading
Also cited by: From $2,000 to $100k: Ross Cameron's 46-Day Day-Trading Playbook
- @smartedgetrading.net https://smartedgetrading.net/blog/how-professional-traders-evaluate-automated-trading-systems
claude · algotrading · tradingview · quantconnect · quantpedia · backtest · stock market