Trading Bear

On my journey to build a portfolio of systematic strategies, I will start by backtesting the following ideas. I will use Codex and Python. Breakout: - Donchian Channel with trailing stops (on Daily and H4) - Intraday breakouts based on ATR (volatility) and key high/low levels (yesterday, ATH, 52‑week high/low, and pivots) Exit criteria are crucial here: I will test both time‑based exits and trailing exits using MA/VWAP. Momentum: - Daily timeframe: trend continuation after a pause - TDI - and MA‑based momentum strategies (expected to fail) - VWAP momentum (expected to fail) Mean Reversion: - Standard deviation bands around VWAP I plan to backtest these strategies in a raw form (without pre‑filtering) as well as with regime filtering, distinguishing between low‑ and high‑volatility periods. Instruments Stocks and crypto in the beginning. I would really appreciate further insights and tips on what to pay particular attention to. I have no experience trading mean‑reversion strategies, as I have always traded breakouts and momentum/trend strategies.

@PKycek Does this mean for smaller, personal accounts the achieveable results would theoretically be better if extend to eg top 100? Or will spreads etc eat the edge there, regardless of size

Most of crypto is not actually tradable. The volatility of the underlying universe sets the upper bound on what a systematic strategy can extract. The top 50 crypto futures sits at 80-100% annualised volatility, roughly six times equities. But volume is what actually sets the constraint on what is actually tradable at scale. Without liquidity, the opportunity is theoretical. A systematic strategy that cannot enter and exit positions at scale is not a tradable strategy for institutions. The chart below shows the share of total perpetual futures dollar volume captured by the largest crypto names, over the last three years. The top 50 crypto future names hold around 90% of all dollar volume traded. The remaining 500+ futures share the other 10%. Most of them are not realistically tradable for institutional systematic strategies. The bid-ask spreads are wide, market depth is thin, and slippage costs eat the edge before it can be captured.

You miss a critical point here - which to be fair is hard to answer without further statistical details. But there is - percentage of time in the market, % of capital used, and not to mention the better risk adjusted returns - which in combination with other strategies could lead to a much higher return on a portfolio level

@SystematicPeter with no benchmark any equity curve is garbage. what if you ask Claude to overlay on it the returns of S&P 500 for the same period. Spare you the trouble. Here it is - S&P in red. S&P did better, no trading fees, no slippage, Zero work.

A single-stock breakout backtest can look great and still be misleading. Take a simple Donchian Channel breakout + trailing stop on TSLA, like I shared yesterday. Strong result? Sure. But picking TSLA first may already be the optimization. You selected the one stock that happened to fit the pattern. A better approach is to test the pattern on a systematically defined, point-in-time universe: * define rules for the universe * rank stocks each day * select the top 5 names * trade the same simple long breakout on that basket * repeat every day Now you are no longer testing a lucky ticker. You are testing a repeatable process with far less hidden discretion and much less survivorship distortion. After fees and slippage, this version shows backtested Sharpe > 2.5 with low correlation to the broad market. That is the kind of simple, systematic edge I am happy to add as a new sleeve to my portfolio.

Trump on Strait of Hormuz.

@SystematicPeter Is the Portfolio with 33% allocation each or 100% depending on signals?

One of the easiest ways to misjudge a trading system is to test it in isolation. I am about to deploy a new slower 4H breakout model. On a single asset class, the results were only ok. Not bad, but not convincing enough on their own. Then I tested the exact same logic with the exact same settings on one equity index, one metal, and one energy market. Completely different picture - even with fees and slippage included. That is a good reminder that tradable vs non-tradable often depends less on the system alone and more on the portfolio context. The reason is simple: an edge is rarely present nonstop across all regimes in one asset class. But when the same edge is picked up across different asset classes in different regimes, the result is often a much stronger system.

@optionsly @quantian1 15 sharpe? Where? 😂

@quantian1 You can get 5-15 sharpe at 50k account value tbh. Lots of low capacity fruit

At a 50k account value, going from a 0 effort 60/40 at 0.5 to developing and deploying a 1.5 sharpe quant strategy is an incremental gain of $5,000 per year. Your time would be more efficiently spent waiting tables.

@AtcoTrader @ConcretumR Wow! Then this really surprises me - as I would assume a lot of noise, but will test myself and check out the distribution

@TradeWithABear @ConcretumR Not on this one!

Here's a very simple system that I trade - opening range breakout on nasdaq. Rules: Entry in the direction of the break of the first 5min bar. Exit either at end of day or at stoploss that is a % of ATR14. CAGR 41,51% Max DD 21,1% 650 trades Idea from @ConcretumR

@PKycek Thanks - I already got the book on my list. I was just curious if you are still trading inside bar/pinbar setups though?

@TradeWithABear Thank you! Some of them are better than others. We also put together a book to explain the basisc of our approaches in a bigger detail... amazon.com/Algorithmic-Cr…

1/ When does it make sense to go MR long after a large selloff? In Top40 Binance perps by volume (2020–2026), the answer is fairly clear: For 5-day drops, the sweet spot is -30% to -50%. That is where mean reversion still works Beyond -50%, these are usually not attractive MR longs anymore.

@yuriymatso sooo you expect one last pump? haha

@SystematicPeter So it‘s for quick „idea testing“?

@TradeWithABear It’s usually much faster.

LLMs can backtest trading ideas surprisingly well - if you feed them clean data. It’s insanely convenient: the model writes a “just enough” backtester for the exact hypothesis you’re testing. But here’s the trap: LLMs sometimes make tiny implementation mistakes. And one small bug can turn a solid idea into total garbage - or total garbage into a “holy grail”. My current workflow that keeps me honest: Step 1: Let the LLM prototype + backtest in its own script (I use Claude Code) Step 2: If results look real, my workflow forces it to re-implement the same logic in a proper framework (I use NautilusTrader). Step 3: Compare outputs - equity curve + trade list must match (or be very close). Step 4: If it matches in Nautilus, odds are the backtest is actually correct. Best part: LLMs can port even complex strategy logic into NautilusTrader without hesitation. Screenshot context: Left = LLM “quick backtester” report Right = same strategy re-coded by the LLM inside NautilusTrader Do you have a validation step like this - or do you trust the first backtest that looks good?

@macrocephalopod Well tbf, couple years back you could do this with Bitcoin…manually…crazy days

Everyone should try this with stocks. I bet you would make a ton of money. Don’t worry if it seems like it doesn’t work at first. Just keep trying.

@TraderOrion Are you using Tradingview Backtesting? I‘m pretty sure that there is some sort of bias or error in this backtests

Ethereum automated trading strategy is now 50% done. Updated Backtest with 0.3% fees and 5 ticks slippage Trades = 106 Pnl = +3330% Win rate longs = 83% Win rate shorts = 81% Sortino ratio = 2.6 Profit factor = 3.35 Buy and hold instead = -16% 2022 = +188% 2023 = +72% 2024 = +112% 2025 = +137%

@KobeissiLetter Stock is up 3% now. If Bulls are lucky this ignites another leg up for the whole market.

BREAKING: Nvidia stock, $NVDA, surges above $200/share after reporting record quarterly revenue of $68.1 billion.

@ripster47 Let‘s just hope for a non-event

$NVDA Earnings 🚨 What happened? Where are you Jensen?

What is extremely noticeable in the last couple years. Gap between Institutions and Retail got smaller. And somehow retail traders are even at advantage. While Instis have to deal with more and more regulatory overhead, the costs of implementing technology + data etc. got so small that almost everyone can run a decent setup at home. It will not stop, just get different I guess. Core principles will always work.

I get asked a lot: “Aren’t you afraid AI will destroy trading? That you’ll have to shut down your systematic vehicle because markets will become too efficient?” Honestly - I don’t know what the world looks like in 10+ years. It will change. A lot. But for small professionals and retail systematic traders, I see AI as a blessing for the next couple of years, not a threat. Here’s why: - Yes, markets may get a bit more efficient. - But inefficiencies won’t disappear - they evolve. - AI won’t magically create alpha. - But it does compress the research cycle - more tests, faster fixes, less grunt work. “But how is it possible that markets won’t become much more efficient if others use the same tools?” Because AI by itself doesn’t generate alpha - it multiplies execution and experience. And I think the majority of retail traders will still try to fight the market “by hand”. So no - I’m not afraid systematic trading will “stop working” because of AI in the coming years. On the contrary: I’m convinced these are the years that can bring substantial profits to smaller systematic traders who are open to using new tools.

@yuriymatso @SystematicPeter Trying to become the textbook example of selection bias? I have another strategy you could backtest: 100% long Bitcoin in 2010 - hold until October 2025. ...what you should do: Get Quality data. Backtest it on HISTORICAL index constituents (eg. S&P500)

@SystematicPeter No look ahead bias. All trades are executed the next day open. So, the returns are absolutely what they are. The strategy is expanding its universe -- it started with 70 stocks initially but I've been adding more and more as they hit my momentum radar.

Seeing a US stock rotational strategy claim +195% annualized with Sharpe 2.92 made me stop and double-check what I was looking at. Published methodology sounds totally “normal”: - composite momentum / multi-factor score - QQQ trend filter - buy top-ranked stocks - rebalance daily So if the edge isn’t exotic… how can the results be this good? First thing I suspect: survivorship bias. Survivorship bias (plain English): - Backtest using today’s universe (current constituents/symbols). - That’s like backtesting 2015 while “knowing” which AI-related stocks would end up being the big winners. - But historically, that universe was messy: losers got delisted, acquired, or went bankrupt - and often vanish from common datasets. - Result: you’re effectively testing only the survivors, and performance can look unreal. Of course it could also be other errors: lookahead / non-point-in-time factor data etc. I just don’t believe performance like this is real without a catch. What’s your base-rate expectation for a clean US momentum rotation like this? Would you trust numbers this high? Stats: systemtrader.co/gemini/perform…

I did some backtests on Crypto and from my experience with them the CME futures make trading them easier (at least for my backtests). My explanation is that the 24/7 market creates a lot of time where there is just „noice“ so signals vanish and edge erases. What is your thought on this?

@Valckrie I use Codex via CLI and on a VPS via Openclaw every day. Mainly for researching/backtesting systematic strategies. And it improved my productivity/speed a lot

Interested in people's experiences with agentic coding tools - have you tried vibe coding and how regular do you use those tools? (Claude/Codex/Cursor)