Vibe Trading

So I have been vibecoding for a while now. Started off building indicators in Tradingview, now at the point where I am monitoring a dashboard of algos.
I've tested 50+ hypotheses and hundreds of strategies that are published on the internet, over 8 years' worth of backtest data. Thankfully found some edge to exploit.


Over the coming months, I will hopefully show the wins and pitfalls of trading my x2 $100k XFA's from Topstep.



I won't be revealing my core strategies fully, but each day, the idea is to test more and see if anything can either compete or beat what I already have deployed. That I can do publicly. I'd be interested to chat with anyone who is already at this stage or wants to exchange ideas on the best way to test trading strats with Claude code etc.

I troll the internet looking for strategies that I can deploy. Here is one that I have adapted to work within the TopStepX XFA accounts. I am forward testing it via paper trading. It is based on a trading concept that Linda Raske had published. Simple continuation/momentum trade. "The Raschke Anti pattern identifies buy-the-dip moments within an established uptrend using Stochastics." It's simple, so coding isn't a problem. Which means it is easy to watch via a Tradingview or MT5 indicator.

Today I built an app. If you investigate how best to use the LLMs it is not to ask them how to come up with a trading strategy; it is for them to process the data of how you trade and how there could be improvements. So, I have made a small little voice-to-text app, which I now use to record my process/observations/trade ideas and trade outcomes. Add some screen grabs, and then let the AI understand what we do.

Great stuff. I am looking to get into vibe coding according to the strategy that i have, and probably exploring other strategies and the advantages they offer. As someone who is experienced, how would you advice that I begin the AI journey?

Don't waste your time trying to trade anyone else's strategy that they post on the internet. Unless you are just going to vibecode a Tradingview / MQL5 indicator and trade discretionary trade setups.

If you have a strategy that you are currently trading, just tell one of the top LLM models what it is you see, how you came up with the trading plan, and just keep feeding it your info and try and make it see what you see. Give it as many rules as possible in terms of what you do on a regular basis, and what you avoid doing. When you feel you have a synergy between what you do on a regular basis and that the Ai can follow along and describe it back to you, then ask Ai to find you as much free stuff as possible. OHLC 1-min data, was the biggest thing I needed for ES and NQ. Then I built out a testing plan, which validates each idea through a series of Gates. If the back test passes all 3 gates, I get to paper trade it.

This is an example of a trading strategy that isn't yet in the live trading environment but is paper trading. Unfortunately for me, Topstep requires all positions are closed at the end of the day. Though for a personal account, this would appear to have been a great entry into a trending move, even though the trade is managed on the 5-minute time frame.

3-Gate Results (NQ/MNQ, 5-min, RTH):
- Gate 1 (significance): PASS — p=0.034, N=408
- Gate 2 (expectancy): PASS — $5.93/trade, $507 max DD
- Gate 3 (walk-forward): PASS — 22/34 windows profitable (65%)
- Robustness: 6/6 ±20% variants all positive

Summary stats: Sharpe 1.67, profit factor 1.28, 54.2% win rate.

I am getting messages asking about how to start Vibecoding, so rather than me repeating myself, here is a little run down of how I would start:

1. Have an edge that you trade, or find someone with one and copy that (no need to re-invent the wheel)
2. Describe what, when, why and which asset you trade to a large LLM (I use claude, but I have used CGPT)
3. Don't ask the LLM to make up a strategy, ask it to copy yours, and make an indicator etc.
4. When you are happy it is doing the basics right, ask it to make a back test over several years of data. I wrote this https://www.rfastrat.com/module1
5. When you have verified that your strat has worked over different times and market conditions, continue to trade it paper then live
6. If point 5 doesn't work over the back test data - go back to the strategy and make an amendment, don't re-do or hop.
7. Rinse-&-Repeat

I am not saying no one should pay for trading journal software, but I am finding it hard to believe that anyone will in the near future. Traders have always been told to keep a journal, to periodically review that journal and then adapt to trading conditions and mental stability. I personally have always had great intentions of keeping a journal, setting up Excel Spreadsheets, buying into Edgewonk, updating the TopstepX journal etc. etc. but when a string of losers come along, I am not so keen to keep up with the admin. But with this system, I am not trading, so if the losses mount up, there is something seriously wrong with the trading logic, and that isn't dependent on whether I got enough sleep last night or had an argument with my wife early that day. I need the Ai to realise that maybe the market conditions we're trading today didn't show up in the backtest data. eg. Trumps Truths, generally quite spiky in the markets, are they hitting my stops?

So I asked my fave LLM to come up with a list of things that any retail trader should be aware of and one of the elements was this:
.

The next task was to implement that sufficiently within the code and then come up with a way to display the findings. The only thing left is a lot more trades to be taken.

Started tracking the details of the trades within a journal with the aim of self-improving the strategies to adapt to different market conditions.
The goal is to also size up with confidence and a lot of that will be determined by the MAE/MFE and Return on Risk. It's going to be a slow grind getting >100 trades across a variety of strategies.

Todays trade hasn't been logged yet, and it was a loser for $88.74, we have about an hour to go before the end of the week and I am not expecting any other trades to fire off to end the week. We're above the original watermark, so happy with that. But mostly happy with the fact that the system has traded without any new bugs appearing. Leaving me to concentrate on improving the data captured, and then what we do with it. Also, have spent quite a lot of time going through strategies on FF, but haven't found any that cross over to mechanical / algo trading. The majority of profitable traders would seemingly be using 99% discretion, loosely based on some trading set-up rules.

Results = Trading 1 MNQ

Scaling with confidence only happens when you stop looking at the dollar amount and start looking at the Standard Deviation of your returns. If you know that your strategy’s worst-case behavior is within a certain range, upping the lot size isn't a gamble it's just a business expansion.
Are you planning to use a dedicated journaling software to run these MAE/MFE calcs, or are you building your own analytical layer to plug into your EA?

Spent the weekend doing a major update on the dashboard, copied over the metrics from TopstepX, so I don't have to switch in and out of their platform to view the same info, and also added a correlation and portfolio matrix. My plan is to have as many orthogonal strategies as I can running at the same time to try and capture as many different regimes as possible, without adding to the risk or overall DD.

Just 2 trades today - was a messy day, considering we had a big drop from the open

Started a new Topstep $100k Combine today. Now that the XFA's are trading, I will slowly add new Combines into the mix and see if I can get the system to eventually trade 5 XFA's. That way, if I win or lose ~$200 per day with 1-micro as per the days so far, I will potentially have +- $1000 days.

I'm also testing out a renko + MACD system with this new Combine so if that works out as expected, I'll attempt to code it into the live trades.

Not expecting much today

This set-up used to be in my portfolio of strategies, but it got superseded by others. It still makes sense though for those looking for a false break of a range
--Update--
The fact that this reversed without hitting TP is why the system dropped the strat.

The First Hundred Days of Running a Systematic Futures Account

Three weeks ago I switched on the first live version of a system I have been building since January. As of today, the account sits at a profit of $642.66 against a $6,000 target, with $2,659.82 of trailing drawdown still in the tank. That is the unglamorous headline. The interesting part is everything behind it.

What the account looks like

The live account is a Topstep Express Funded V2. Profit target $6,000. Trailing drawdown $3,000. Current profit from breakeven is $642.66, which is 10.71% of the way to payout. The account has never been locked, never blown, and the equity curve has the shape you would expect from a research-driven system in its first few weeks: a couple of decent days, a couple of painful ones, and a slow rebuild.
The peak equity so far was $340.18 on April 20. Two stops on H46 the following day cut that back to negative territory before a run of manual discretionary trades rebuilt the balance to where it sits now.

The research graveyard

The real work of the last few months has not been trading. It has been killing ideas. The project has tested 53 hypotheses. Most are dead. The handful that made it to live trading are the small minority that survived a three-gate filter: p < 0.05 on 100+ trades, positive expectancy after realistic costs, and robust across plus or minus 20% parameter perturbation.
In the last two weeks alone, the following were killed:

1. Three Tom v2 strategies (H62 1BN fade, H63 inside-bar SRS, H64 FOMC bracket). All negative across ES and NQ despite strong theoretical priors. H63 was especially interesting because its filter actively made the strategy worse, the opposite of what its author claimed.
2. Four trend and momentum hunts (H66 through H69). The entire trend-continuation family is now considered exhausted for ES and NQ, which is itself a valuable negative result.
3. An IRONCLAD daily/15-minute SMC template (H70) whose author claimed 957% on forex. Tested on ES and NQ, it was negative on one and weakly positive but catastrophic in drawdown on the other. The 957% figure turned out to be the best of 8,700 parameter combinations, a near-certain p-hack.

The lesson I keep relearning is that charisma and production value do not predict edge. A polished PDF with Pine indicators is usually just the best-of-thousands cherry. Real edges survive being attacked from every angle you can throw at them, including deliberately pointing the backtester at the parts of the data the author never showed.

What is actually live

The active portfolio on the Topstep account is small on purpose:

1. H46 DR CE Counter (MNQ, primary). Twenty live trades so far, 50% win rate, net -$238.06. Backtest expectation is $13.57 per trade. Twenty trades gives an expected P&L around $271. The live deviation is roughly one standard deviation below expectation. Uncomfortable, but within the cold-cluster envelope the walk-forward analysis predicted.
2. H48e bias-candle pullback (MNQ). Three trades, +$96.78. Early days.
3. H60 squeeze-breakout (MNQ). Running in shadow mode to accumulate a live-versus-backtest sample before committing capital.
4. H54 swing pullback and Anti Long NQ. Paper-trial registered. Awaiting first signals inside a 12-week observation window.
5. H37v2 ES queued to switch on once NQ trading is stable.

Alongside the algos I have been running a manual discretionary overlay. Over the last fortnight those trades added $613.20, which is most of the current account balance. This is deliberate. The macro regime in April has been noisy, with overlapping POTUS, Iran, and inflation drivers. H46 is going through a normal cold patch. I would rather bridge drawdown with discretionary trading I have ten years of experience executing than watch the machine bleed while waiting for its mean to reassert.

The numbers nobody shows you

Here is the April equity trail for the current account, algo only:

1. April 9: +$248.78
2. April 10: -$303.46
3. April 15: +$242.28
4. April 16: -$17.46
5. April 17: -$38.98
6. April 20: +$208.02 (peak HWM)
7. April 21: -$385.98 (two H46 stops in a single session)
8. April 22: +$56.26
9. April 23: $0 algo, discretionary still running

A peak of $340.18, a single painful session that nearly wiped the gains, and then a rebuild driven by manual trades. Every systematic trader presenting a straight-line equity curve is either lying or selectively showing you their best six months. Real curves look like this. The question is not whether the drawdowns come. The question is whether you know, statistically, how bad they are allowed to get before the system itself should be suspected.

Why I am doing it this way

The goal is not to hit the $6,000 payout as fast as possible. It is to build a system that can pass Topstep combines repeatedly, with a statistical signature that matches what the backtests predicted. A first payout matters less than the distribution of outcomes across the next hundred accounts.
That is why every strategy has to clear three independent gates before any live capital touches it. It is why the H62, H63, H64 kill is more valuable than the H48e lift. It is why I am willing to run discretionary trades alongside the algo while the algo accumulates the sample it needs.

What is next

1. Clean signal days on H46 to restore the algo-only equity curve
2. First live fills on Anti Long and H54 to widen portfolio coverage
3. Continuing to hunt for uncorrelated edges, specifically not more variants of the strategies that already work (the trend family is exhausted, so the next hunts need to be structurally different)
4. Scaling H46 to two micro contracts after 50 live trades land inside the backtest confidence interval

This is the honest status report. A small profit on a small target. A handful of strategies surviving, many dead. A machine that is running, a trader still actively involved in bridging it, and research continuing in the background.
The measure of the work is not the $642.66 sitting in the account right now. It is the discipline of the filter that put only five strategies in front of the market and killed forty-eight others before they could do damage.
That filter is the product.

I just did a deep dive into Volume Profile


This is for one of my strategies as I considered using VP as a filter