Diary of a Pro Quant | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Is random really random? 231 replies

Pro Traders in a Random Market 60 replies

Walking the walk 2 replies

Alpari Direct Pro vs Alpari MT4 Pro 4 replies

Trading Diary/Journal 5 replies

Attachments: Diary of a Pro Quant
Exit Attachments

Diary of a Pro Quant

  • #1
  • First Post: Edited 12:40pm Jun 1, 2023 8:20am | Edited 12:40pm
Welcome to my thread. I'm a professional quant and I'll use this thread to share some of my forex research, discuss some technical aspects of what I do, and document my own progress as I explore a variety of areas of FX.

Unless mentioned otherwise, I'll be using Python exclusively. I'll generally try to keep this at a high level as many aspects of what I do are broadly applicable to traders, whether they use Python or not. What technical things I do mention will be along the lines of, "I was looking to do this, I tried x, y, and z libraries and found that z offered the perfect solution."

This thread is also for you, the reader. Have an interesting idea? Share it here! Have a question? Ask away! I enjoy collaborating with other experienced traders and I'm always happy to help out newer traders so please don't hesitate to chime in.

You're also welcome to DM, or reach out to me on Twitter here: https://twitter.com/ForexStats
  • #2
  • Jun 1, 2023 1:09pm Jun 1, 2023 1:09pm
I started trading a long time ago and over the years my approach towards the markets has evolved. Now, I do things very differently from when I started and how most traders operate.

  1. I rely 100% on statistics to guide my decisions, but my statistical edge always has to be supported by an underlying economic rational to make it trustworthy... Subjective chart patterns and technical analysis serve no purpose. If it's worthwhile, program it and let's see your Sharpe.
  2. I don't use platforms anymore. I'll use ForexFactory for quick quotes, but otherwise I'll use my Python console for everything else. At this point, it normally doesn't even occur to me to use something like TradingView even for simple things like viewing a chart of EUR/USD for the past year. I go to my IDE, pull data for EUR/USD, adjust the date range, and plot it.
  3. I HATE opinions. Guessing where the market is going to go due to untested theories is a complete waste of time. This applies both to traders on here and FX Analysts you see on Bloomberg. Show me your backtest, give me your data, but please keep your subjective opinion on the markets to yourself... To be clear, discussing a market or economic relationship can be very useful because I can independently test the idea. But, even well supported conjecture about the markets just isn't good enough.
  4. I trade lower frequency. Daytrading is for market makers, not retail investors. In general, the more frequently you trade the lower your odds of success.

  • #3
  • Jun 2, 2023 3:17am Jun 2, 2023 3:17am
Would love to see the results of your trading! By statistics, does it mean that you use mathematical models to forecast the direction of a currency pair?
  • #4
  • Jun 2, 2023 5:15am Jun 2, 2023 5:15am
What resources would you recommend for someone who wants to become a pro quant?
  • #5
  • Jun 3, 2023 8:54am Jun 3, 2023 8:54am
Quoting BlackCap
Would love to see the results of your trading! By statistics, does it mean that you use mathematical models to forecast the direction of a currency pair?
Yes, I use models to trade but I'm specifically referring to performance statistics of a strategy like those you'd find on a Trade Explorer.

Quoting Kefada
What resources would you recommend for someone who wants to become a pro quant?
A "quant" is a very general term and the resources you need depend on your specific goals. Derivatives pricing is one route, HFT another, and so on and so forth. A strong coding ability is a universal qualification. Beyond that, it just depends on your specialization.
  • #6
  • Jun 3, 2023 9:07am Jun 3, 2023 9:07am
Quoting ForexStats
{quote} Yes, I use models to trade but I'm specifically referring to performance statistics of a strategy like those you'd find on a Trade Explorer. {quote}
Are the models computed from released financial data?
  • #7
  • Jun 3, 2023 12:59pm Jun 3, 2023 12:59pm
Quoting ForexStats
I started trading a long time ago and over the years my approach towards the markets has evolved. Now, I do things very differently from when I started and how most traders operate. I rely 100% on statistics to guide my decisions, but my statistical edge always has to be supported by an underlying economic rational to make it trustworthy... Subjective chart patterns and technical analysis serve no purpose. If it's worthwhile, program it and let's see your Sharpe. I don't use platforms anymore. I'll use ForexFactory for quick quotes, but otherwise...
Perfect. I guessing unless you don't know where the price will be 100%.

I hate guessing unless you don't know where the price will be 100%. But statistics isn't answer...

So being convinced that price will be a zone is an opinion?
KEYWORD is C O N V I N C E D...
  • #8
  • Jun 4, 2023 3:14pm Jun 4, 2023 3:14pm
Quoting BlackCap
{quote} Are the models computed from released financial data?
Generally yes, alt data can be very valuable when you have access though.

Quoting Ihlas
{quote} Perfect. I guessing unless you don't know where the price will be 100%. I hate guessing unless you don't know where the price will be 100%. But statistics isn't answer... So being convinced that price will be a zone is an opinion? KEYWORD is C O N V I N C E D...
Knowing where the price will be 100% of the time is impossible. Attempting to trade with perfection implies you have a deterministic view of the markets, which is a mistake I made early in my career. I always emphasize the importance of risk management because many aspects of trading are simply unknowable. A good model counts for a lot, but position sizing and risk management are two very important and generally under appreciated areas of trading. Rather than striving for the "holy grail" traders are better served finding a decent strategy, then optimizing the position sizing / stop losses / etc.
  • #9
  • Edited 11:35am Jun 6, 2023 11:20am | Edited 11:35am
I developed a new chart showing the seasonal cycle. It can be used on any asset so let me know if you have any requests, now or in the future.
Attached Image (click to enlarge)
Click to Enlarge Name: SPX.png Size: 20 KB


This is how it works:

  1. Calculates the log return for every day start to finish.
  2. Break down every year into trading days (1-255ish, with 1 being the start of the year and 255 being the last trading day of the year).
  3. Group all the trading days together and calculate the mean return. (ie mean return of all day 1s, day 2s, etc.)
  4. The cumulative sum is taken and multiplied by 100.


There are a few different ways this could have been done, and I may make changes later, but right now I think it works.

  • #10
  • Jun 6, 2023 11:37am Jun 6, 2023 11:37am
Quoting ForexStats
I developed a new chart showing the seasonal cycle. It can be used on any asset so let me know if you have any requests, now or in the future. {image} This is how it works: Calculates the log return for every day start to finish. Break down every year into trading days (1-255ish, with 1 being the start of the year and 255 being the last trading day of the year). Group all the trading days together and calculate the mean return. (ie mean return of all day 1s, day 2s, etc.) The cumulative sum is taken and multiplied by 100. There are a few different...
Could you add error bars? It would be nice to see the deviations from the mean.
To free Gazans of Hamas, use whatever it takes.
  • #11
  • Edited 8:05am Jun 7, 2023 5:17am | Edited 8:05am
Subscribed! Looking forward for new fresh ideas.

Some thoughts on your seasonal cycle idea above:

* These kinds of a statistics are very sensitive to outliers. One day with an exceptionally large profit/loss can skew your stats. Instead of the mean return, I recommend using the median return as this is much less sensitive to outliers.
* To see is this idea actually works, try coding a walk-forward simulation. For example, use a moving window of the last 50 years. Calculate the stats over these years and use the stats to predict the returns of the next 250 trading days. Then move the window 1 year and do it again. Go long for the day if the median return of the last 50 years exceeds the trading cost, else remain flat. That way, try simulating 1977-2023 and see what equity curve you get and compare that against a simple buy-and-hold strategy from 1977 to 2023.
* Don't use D1 bars for your calculation! Why? Because that would imply you need to enter/exit the market at midnight when the spreads are at maximum. Instead, calculate your own daily returns at, for example, 04:00 AM.
  • #12
  • Jun 7, 2023 8:04am Jun 7, 2023 8:04am
I couldn't resist and coded the above out of curiosity. Here is a walk-forward simulation from 1977 to 2023 comparing daily trading based on median return versus a buy-and-hold since 1977:
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 19 KB


Unfortunately, the model does not seem to provide an edge.

Perhaps the noise level of daily returns is too much and instead we should consider the return over a multiple days?
  • #13
  • Jun 7, 2023 8:18am Jun 7, 2023 8:18am
Quoting ForexStats
I developed a new chart showing the seasonal cycle. It can be used on any asset so let me know if you have any requests, now or in the future. {image} This is how it works: Calculates the log return for every day start to finish. Break down every year into trading days (1-255ish, with 1 being the start of the year and 255 being the last trading day of the year). Group all the trading days together and calculate the mean return. (ie mean return of all day 1s, day 2s, etc.) The cumulative sum is taken and multiplied by 100. There are a few different...
Hi really interesting. Could you do it with gold please ?
Best
  • #14
  • Jun 7, 2023 8:53am Jun 7, 2023 8:53am
Quoting DrDave
{quote} Could you add error bars? It would be nice to see the deviations from the mean.
I don't believe so in this case, if I understand correctly. Being that this is the cumulative sum of log returns, the error at the beginning would be relatively small and huge at the end.
  • #15
  • Edited 9:06am Jun 7, 2023 8:53am | Edited 9:06am
Quoting yoriz
Subscribed!
Thank you!

Quoting yoriz
* These kinds of a statistics are very sensitive to outliers. One day with an exceptionally large profit/loss can skew your stats. Instead of the mean return, I recommend using the median return as this is much less sensitive to outliers.
You make a good point, I was debating between mean and median myself. Here's a chart showing mean vs median for the S&P 500:
Attached Image (click to enlarge)
Click to Enlarge Name: meanvsmedian.png Size: 21 KB


After reading your post I initially thought median was more representative of what to expect going forward, but now that I've seen the chart I'm not so sure. It's basically just a 45 degree line going up, unsurprising of something with an upside bias like an equity index.


Quoting yoriz
* To see is this idea actually works, try coding a walk-forward simulation. For example, use a moving window of the last 50 years. Calculate the stats over these years and use the stats to predict the returns of the next 250 trading days. Then move the window 1 year and do it again. Go long for the day if the median return of the last 50 years exceeds the trading cost, else remain flat. That way, try simulating 1977-2023 and see what equity curve you get and compare that against a simple buy-and-hold strategy from 1977 to 2023.
My plan is to add in different cycles next. This is just the yearly cycle, but it won't be hard to get average/median returns for the presidential cycle, decennial cycle, etc. Or filter by years when the Fed is hiking, or when the unemployment rate is below the long-term average. Anything that seems relevant really.

Ultimately I plan to produce various cycle charts for every currency, then do a Year-To-Date correlation showing which are/aren't working. This presents some issues too, like what do you do at the beginning of the year when there's no basis for comparison, but they'll be addressed in due time.

I'm currently working on an ML project so it may be a little while before I get started on cycle charts 2.0.

Quoting yoriz
* Don't use D1 bars for your calculation! Why? Because that would imply you need to enter/exit the market at midnight when the spreads are at maximum. Instead, calculate your own daily returns at, for example, 04:00 AM.
Good point, I need to double-check how Yahoo! is quoting this data. They're only providing Daily, Weekly, and Monthly data so I may need to go elsewhere if this presents a problem.

Quoting yoriz
I couldn't resist and coded the above out of curiosity. Here is a walk-forward simulation from 1977 to 2023 comparing daily trading based on median return versus a buy-and-hold since 1977: {image} Unfortunately, the model does not seem to provide an edge. Perhaps the noise level of daily returns is too much and instead we should consider the return over a multiple days?
Nice work! Did you use Python for this?

I'm actually not surprised the model didn't provide an edge. As mentioned above, tweaking this by filtering for more relevant market conditions makes sense. It may take this from being a seasonal chart to more of an historical analog chart.
  • #16
  • Jun 7, 2023 8:58am Jun 7, 2023 8:58am
Quoting rosalieone
{quote} Hi really interesting. Could you do it with gold please ? Best
Sure, here's gold as represented by the GLD ETF.

Attached Image (click to enlarge)
Click to Enlarge Name: GLD.png Size: 25 KB


Gold and other commodities are tricky because Yahoo! Finance doesn't provide historical data for them. GLD works for our purposes, although it's a relatively short history unfortunately... I'll have to try this again using a different data provider later, maybe the St. Louis Fed or Stooq.
  • #17
  • Jun 7, 2023 9:00am Jun 7, 2023 9:00am
Quoting DrDave
Could you add error bars? It would be nice to see the deviations from the mean.
Below are the 2%, 16%, 50%, 84% and 98% relative returns per day of year (aka -2 to +2 sigma):
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 65 KB


I also made a cumulative plot to mimic @ForexStats graph, but it looks very different:
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 14 KB


@ForexStats: How did you match days? I tried these two methods:

  1. Match the exact calendar day (e.g. May 13, 1928 + May 13, 1929 + May 13, 1930 + ...)
  2. Match the 1st Monday of the year (e.g. Jan 2, 1928 + Jan 7, 1929 + Jan 6, 1930 + ...)


But both give different results than your graph. Can you please elaborate how you grouped dates?

  • #18
  • Edited 9:30am Jun 7, 2023 9:14am | Edited 9:30am
Quoting yoriz
{quote} Below are the 2%, 16%, 50%, 84% and 98% relative returns per day of year (aka -2 to +2 sigma): {image} I also made a cumulative plot to mimic @ForexStats graph, but it looks very different: {image} @ForexStats: How did you match days? I tried these two methods: Match the exact calendar day (e.g. May 13, 1928 + May 13, 1929 + May 13, 1930 + ...) Match the 1st Monday of the year (e.g. Jan 2, 1928 + Jan 7, 1929 + Jan 6, 1930 + ...) But both give different results than your graph. Can you please elaborate how you grouped dates?
Your cumulative return looks almost the same as mine when I used median. One slight difference is yours dips at the very end, and mine doesn't because I'm excluding the last two trading days. The reason for that is some years have more days than others, and at the end of the year the sample gets very small so I thought it was best to exclude them.

To group trading days, I split every year into a separate dataframe, counted the trading days, merged them into a new dataframe, then grouped by trading day and calculated the mean/median. I considered just doing it by day of the year, which would have been a lot easier, but in the end I thought it wouldn't work as well. Since there are only 250ish trading days, spreading the year out to 365 would have lowered the sample size for any given day so going by day of the year may have caused problems. Also, being that this is cumulative return, a longer trading year would exaggerate gains/losses in assets with an upside or downside bias like the S&P 500.
  • #19
  • Edited 9:56am Jun 7, 2023 9:29am | Edited 9:56am
Quoting ForexStats
After reading your post I initially thought median was more representative of what to expect going forward, but now that I've seen the chart I'm not so sure. It's basically just a 45 degree line going up, unsurprising of something with an upside bias like an equity index.
Ah, perhaps that explains the difference between mine and your cumulative returns graphs.

Quoting ForexStats
Good point, I need to double-check how Yahoo! is quoting this data. They're only providing Daily, Weekly, and Monthly data so I may need to go elsewhere if this presents a problem.
Come to think of it, if you enter/exit at 23:45 PM when spreads are still ok you will probably get almost the same results since volatility is usually very low at night so not a lot of movement in those last 15 minutes.

Quoting ForexStats
Did you use Python for this?
No, I used Java (and Apple Numbers for the plots). For these simple experiments probably any language will do, although admittedly Python is very convenient for this kind of scripting. However, I do most of my ML using proprietary algorithms that are not available in the Python libraries. Implementing them in Python would be too slow and implementing them in C as a loadable Python library is too much hassle for my taste. Hence I do mostly work with Java which nowadays is as fast as C++ (without the complexities of writing robust C++ code).

Quoting ForexStats
I'm actually not surprised the model didn't provide an edge.
If it were this simple everybody would be a millionaire... ;-)
  • #20
  • Jun 7, 2023 9:44am Jun 7, 2023 9:44am
Quoting ForexStats
To group trading days, I split every year into a separate dataframe, counted the trading days, merged them into a new dataframe, then grouped by trading day and calculated the mean/median.
Ah, I see. That is different from both methods I tried.

Since we are looking at seasonal effects on a daily basis, it is very important the matching is exact. However, here it matters what seasonal event we are trying to find. Certain holidays are always at the same calendar date (and therefore not on the same day of the week). Other events are always on the same day of the week but on different calendar dates (e.g. returning to office on Monday after holiday, anticipating an upcoming holiday on Friday, etc.). So it is hard to tell which way of matching data points is best.
0 traders viewing now