ForumSections

Success Stories in Algorithmic Trading - we need more of them! - page 3

New comment
Oleksandr Medviediev #:
Thanks for the tip, Conor. What's your story?

I managed to pass the FTMO challenges with an EA that used the impulse macd (and standard risk management). My strategy wasn't to leave the EA on the chart, it was to place it on the chart at key times - for example, I needed a short entry on EURUSD when the Fed raised interest rates, so I placed the EA on the chart when it was too late to signal a macd buy, and then it was forced to look for the next macd sell signal (which was the desired situation). I had the macd looking for opportunities on the M15 timeframe, and I use a slow length ma period of 50 - 70 to push past small liquidity corrections that can interfere with the trend on this timeframe.

Thanks Conor. So you pick the moment, the EA does the rest, correct? Was passing FTMO easy with this approach?

The realistic-expectations point is the one most people skip. The traders who last are usually the ones who stop measuring success by win rate and start measuring it by whether the edge survives costs. A system that wins 60% of trades can still lose money if commission and swap eat the winners. The tools that help are the unglamorous ones: cost auditors, drawdown monitors, and honest backtest verification. The miracle EA does not exist, but a mediocre system with good risk control and honest accounting can survive long enough to compound.

Oleksandr Medviediev #:
Thanks Conor. So you pick the moment, the EA does the rest, correct? Was passing FTMO easy with this approach?
Yes the EA does the rest. Passing the FTMO challenge was definitely easier this way, but I didn't trust things at first. I failed previous challenges, and after journaling the trades, it was failed because of my own manual trades that time.
Jevon Naldo Yoshield #:

The realistic-expectations point is the one most people skip. The traders who last are usually the ones who stop measuring success by win rate and start measuring it by whether the edge survives costs. A system that wins 60% of trades can still lose money if commission and swap eat the winners. The tools that help are the unglamorous ones: cost auditors, drawdown monitors, and honest backtest verification. The miracle EA does not exist, but a mediocre system with good risk control and honest accounting can survive long enough to compound.

Thanks Jevon. Did cost auditing change anything in your own results?

What’s your story?

Jevon Naldo Yoshield #:

The realistic-expectations point is the one most people skip. The traders who last are usually the ones who stop measuring success by win rate and start measuring it by whether the edge survives costs. A system that wins 60% of trades can still lose money if commission and swap eat the winners. The tools that help are the unglamorous ones: cost auditors, drawdown monitors, and honest backtest verification. The miracle EA does not exist, but a mediocre system with good risk control and honest accounting can survive long enough to compound.

Well said. Here's a related excerpt from my user Profile:

Here's a little known position sizing method to help with scaling up.

Let's assume that you have a simple entry and exit EA with a profit factor of 1.5, for example:

1. Run it through the Tester. Testing on the Real Ticks of your live broker-dealer's data is ideal if your machine's CPU and RAM can handle that. I prefer to run through 4000 round trades or more.
2. Read the Tester Report carefully. Additionally, use my custom ProfitByHourReportService (in my Blog, below) to analyze hourly profit by trade entries─not by trade exits as the native MT5 Tester Report does.
3. Pay particular attention to the months, days, and hours in which the EA is profitable.
4. Code a time filter that blocks entries during the worst times in the EA. See my millisecond-based "Trading Time Filter for Any MT5 Expert Advisor" code snippet (in my Blog, below).
5. Run a fresh backtest with the time filter implemented.
6. Examine the average count of consecutive loss trades─let's call this "NL".
7. Code automatic position sizing based on say, 1% of account balance, into the EA.
8. Code an exception to that 1% position size that triggers when NL+1 occurs (the EA has exceeded its average loss count).
9. Code an overriding minimum position size of say, 0.01, to operate when that exception occurs.
10. Code a resumption of 1% position sizing that triggers when one trade closes in profit.

Of course, additional backtests should be run as necessary following each code edit. The purpose of the process outlined above is to solve the mystery of attempting to predict when the EA's logic will be profitable versus unprofitable.
To be clear, I'm presently using this in my live algorithmic trading.

What I believed that was wrong was that my own backtest engine measured the market. For months my index breakout setups tested positive in a bar based Python engine I wrote, and I kept testing filters on top of those numbers. Then I ran the exact same engine on a fake market, every one minute bar mirrored around its open with probability one half, so there is no direction left to find. It still booked a profit on that coin flip market in all 34 setup and market cells I ran, and for some setups almost the whole edge I had measured was already there on random data. The cause was two small fill assumptions. Entries were booked at the level although the signal only confirmed a bit later, and the trailing stop was filled at its level on the next bar even when the market had already opened beyond it. After fixing that, none of the setups I was trading live showed an advantage over random. The habit I kept is to run every test once on data where the answer has to be zero before I believe the real run.

Great story, Timon. Can you explain the fake market test in 2-3 simple sentences? Many readers here, me included, are not native English speakers.

Oleksandr, sure. I took real price history and removed its direction: every one minute bar was either flipped upside down around its open or left as it was, chosen at random like a coin toss. On data like that no strategy can have an edge, so an honest backtest has to show about zero profit before costs. Mine still showed a profit, which told me the profit came from mistakes in how my test filled the orders.

123
New comment