Machine Learning Forex System (MLFX)

Welcome to all!

I'm Drew, one of the co-founders of MLFX a Forex algorithm based on 10 years of university-level research by my business partner and friend Amaury Hernandez. Our algorithm (Hermes) uses multi-agent systems, intuitionistic fuzzy logic, and bio-inspired machine learning to predict and trade the Forex markets. If that made you scratch your head (trust me you're not alone) then let me break it down for you.

essentially each "AGENT" can be thought of as an expert human trader
These agents each have a set number of unique INPUTS (such as RSI)
Those inputs have RULES assigned to them (if RSI is above 80 do xyz)
a GROUP of RULES is essentially a trading strategy
fuzzy logic lets agents MEASURE HOW TRUE/FALSE each rule is, as opposed to binary true/false
The machine learning aspect iterates through thousands of possible combinations looking for optimally configured agents

If you're still struggling, think Quantopian but instead of human quants creating algorithmic trading strategies, machine learning is creating those strategies instead.

June 2021 Update:
We've decided to shift our focus onto creating something lots of people can use right now, the problem with what we were previously working towards is that it was uncertain when we would have a product that was actually useful to people. Yes we had metrics in mind that we were tracking, but it could have taken us years to get there and we need revenue much sooner than that. So, what are we working on now?

MLFX MVP (minimum viable product):
We're creating a library of popular/requested Forex strategies that can be easily compared and ranked against each other based on historical and live performance. Strategies become inputs for our machine learning agents who iteratively find better configurations of that strategy.

As an MLFX user, you can follow any of these strategies, plot historical and live trades on our charts, and receive live notifications when new signals are created.

WHY?
Reduce time spent scanning charts for setups that meet your criteria
Giving you more time to focus on trade execution and which setups to trade
Resulting in less trader fatigue and better trading decisions

Reduce uncertainty in the performance of your strategy
Giving you the confidence to stick with a winning strategy through a rough period
Resulting in more consistent long-term trading results

We're currently looking for an experienced *profitable* trader to join our team in an advisory role - You would be assisting with the creation of our products to make them more trader friendly and advising us on areas we should priorities when it comes to algorithm development. For instance, indicator X should be focussed on before indicator Y. Send me a message if you're interested.

Our Telegram group is a great way to keep in contact with us: https://t.me/joinchat/IFyZTEw_uhEHe0Oo5uSAVw

REQUEST A STRATEGY IN THIS THREAD:
If you have a strategy that you want us to add to MLFX then request it in this thread and we'll let you know when it has been added. This will eventually be a paid feature, something like $50-$100 per strategy, but as long as you're seeing this then it's free!

MLFX RESEARCH:
All research found here:
Using Fuzzy Inference Systems for Creation of Forex Market Predictive Models

MONTHLY UPDATE ARTICLES:
July 2020: | Launch of Analysis Page
August 2020 | Implementation of Multi-Objective Algorithm
September 2020 | A Pattern Thingy
October 2020 | 200 THOUSAND Times Faster
November 2020 | Storing and Plotting Data
December 2020 | Manual Dataset Selection & The Mactavator
January 2021 | Bug Squashing Time
February 2021 | The Team Grows
March 2021 | Nested Signals
April 2021 | A Beautiful Vue
May 2021 | Eating Humble Pie

The results for last week were really bad! But we know it's due to some changes that were implemented with the release of our pattern recognition system.
The pattern recognition system is used in the agent creation process to makes sure that agents are created and optimized for similar market conditions. We were also using it to identify which agent pool should be used when creating the signal, which is part of the reason our results have suffered over the last few weeks.
(Above is a look at how our agent pools are currently segmented for each pairing, we can have as many as we want)

We're now using the most activated agent globally (all agent pools for that pairing) to produce the signal.
I'd estimate that we will see about a 20-30% improvement in performance, which will still result in an unprofitable system as you will see in the results below.
But that's fine, the improvements that will make it profitable are coming very soon.

WEEK 7 RESULTS

GBPUSD
W32/L80
PIP = W1397.8/L1385.8
RETURN = +1.22%
AVG. WIN R/R = 2.59

EURJPY
W36/L76
PIP = W1499.8/L1001.5
RETURN = +36.7%
AVG. WIN R/R = 2.98

AUDUSD
W17/L93
PIP = W520.8/L1151.5
RETURN = -38.6%
AVG. WIN R/R = 2.59

USDCAD
W6/L107
PIP = W213.4/L1476.9
RETURN = -61.1%
AVG. WIN R/R = 2.05

EURGBP
W36/L77
PIP = W786.1/L619.3
RETURN = +4.9%
AVG. WIN R/R = 2.26

USDJPY
W8/L104
PIP = W188.2/L1185.9
RETURN = -58.3%
AVG. WIN R/R = 2.00

USDCNH
W20/L93
PIP = W2142.9/L3967.9
RETURN = -14.9%
AVG. WIN R/R = 2.93

EURUSD
W34/L79
PIP = W1369/L1173.3
RETURN = -14.9%
AVG. WIN R/R = 2.36

USDCHF
W42/L770
PIP = W938.3/L637.8
RETURN = +37.7%
AVG. WIN R/R = 2.42

TOTAL RESULTS
W231/L779
PIP = W8862.5/L12635.9
RETURN = -87.9%
AVG. WIN R/R = 2.48

OCTOBER RESULTS RECAP

Week 1

Week 2

Week 3

Week 4

Overview:
Yes, we're aware that the results are negative. This was to be fully expected, we're still testing and tweaking things.
We are actually happy with the results, believe it or not, you can see a dramatic improvement from week 1&2 to week 3&4 which was when we made the following changes.

1. Increased the Training & Creation dataset sizes
2. Switched back to using the most activated agent globally instead of using pattern recognition to identify which agent pool to use, thus being constrained to that agent pool.
3. Started forcing each agent to test the market every data point (every hour)

What we're trying to do is make 1 change that affects algorithm performance and then give it two weeks to verify that the changes are positive. Unless the results are seriously out of whack... in which case we might revert after 1 week.

The next changes we're testing (as of today) are:

1. Storing rates so we can use a larger total dataset, we're using 6,000 data points (6,000 hours) instead of 1,500. This was needed because in some cases the system was struggling to find 2 bullish/bearish/stagnant markets that met the required criteria.
2. Added a rule so that buying/selling agents can only dominate and replace their counterparts.
3. Some performance improvements that will get the speed to where it needs to be. This means we should hopefully have roughly the same number of trades to compare each week.

OCTOBER 2020 UPDATE ARTICLE.
Although we're having to slow down our changes to the algorithm so we're able to verify changes adequately, there was still plenty of development for MLFX in October. Read about all our improvements in this months update article: https://www.mlforex.com/blog/update-...0-times-faster

HUMAN SELECTION OF DATASETS

We've been toying with different ideas on how to best select and rate the datasets used for creating and training our agents for some time now. Currently, we use what we've dubbed the "Pattern Thingy" which is doing and OK job, but can definitely be improved. One of the first steps for us was to identify how the current system is rating the datasets and what they looked like. To do this we have plotted each dataset so it can be viewed from the new /Agents page, as shown below.

Attached Image (click to enlarge)

The Agents page.

Attached Image (click to enlarge)

A plot of the creation dataset for the Bullish AUD/USD Agent Pool.

We are going to display the rating below the plot for each dataset and have a button to export the raw data. This means we can import the data into spreadsheets and try different methods of rating the dataset. Something that will hopefully produce a better method than the currently lacking one. We will also use something that looks similar to the above plot to select the datasets to use, I will manually slice the data, choosing the start and endpoint for all datasets we'll use. Then it will be fed to the Pattern Thingy to rate and sort. Eventually, we could say "we want 10 different pattern categories ranging from Super Bullish to Super Bearish..." and we can then create agents for each of those categories. We're hoping to have most of this implemented within the next 1-2 weeks.

Your results and improvements seem to trend to the better...keep going. The transition to Week 3 from Week 2 was dramatic.

Es

NOVEMBER RESULTS RECAP

I've neglected this thread a bit, no real excuse! Below will be a bunch of posts to catch people up.
It is worth noting that as I write this lots has changed and we are on the cusp of potentially launching our product with some pending changes.

Week 5 | 2nd - 7th
We started off the month with a positive note, our first positive compounded return. Not a huge HOOOORAAAHHH but still a nice milestone. For week 5 we were having performance (speed) issues which resulted in far fewer signals being produced. Week 5 was the first week of results since we increased the total datapoints being used to create our datasets, and we also implemented a rule that stops buying agents from dominating selling agents and the opposite. BUT... Week 5 was the US election, so we didn’t use them to verify our changes as the markets were very volatile. This means we used week 6 and 7 to verify these changes properly.


Week 6 | 9th - 14th
26% compounded return was a good improvement and went towards verifying the changes listed above. if we're able to verify this with another week around that number, we can move forward with the next changes.


Week 7 | 16th - 21st
Pretty bad results for week 7, the markets were very up and down. We moved forward with some new changes to the algorithm. Starting week 9 we will be testing a change to our fitnesses, (what the algorithm optimizes for, think objectives,) from Avg. Revenue, Wins & Losses to Total Return and Avg. Return. By return, we mean the % return made. We're not expecting a major increase in accuracy, rather a decent increase in the risk/reward, time will tell.


Week 8 | 23rd - 28th
Week 8 still had 4 trades outstanding so I called them losers even though they are looking like winners. We didn't have a great number of trades for the week, so although it's a positive week, we're taking that with a grain of salt. Regardless, things are getting better every week which is what we're looking for and it was enough to move forward with the next changes which were to the our algorithm fitnesses or objectives detailed earlier in the article. Spoiler alert… because of this we had our first “OFFICIAL” positive week for week 9, so keep a lookout for that in next months update.


Market Ready Update
It appears that we are rapidly moving towards a market-ready product. I’m confident that we will be officially launching in either January or February 2021. As I write these final few sections the algorithm is fighting with our break-even line of 25% accuracy and it seems to be winning. It’s looking like a consistent 30% accuracy with a 1:3 Avg. R/R is likely what we will launch with. However, two changes being made this week could skyrocket us past that, so time will tell. Just for reference, that accuracy and R/R would lead to +20% Return from 100 trades.

NOVEMBER UPDATE ARTICLE
We made a bunch of improvements during November including the ability to store price data internally and the plotting of that data using plotly.com, there was also a change in fitnesses and major performance improvements. The full article can be found here: https://www.mlforex.com/blog/update-...trading-begins

DECEMBER RESULTS RECAP

Week 9 | 30th Nov - 5th December
Our first “officially” profitable month came as an early Christmas present, and what a result it was! +160% in total and more importantly an avg. of +0.19% per trade. Obviously, we were excited with the result, yet sceptical because it is only 1 week. That being said, our changes are definitely leading us in the right direction.


Week 10 | 7th - 12th December
No results this week as we were making some changes to the system. We added manual dataset selection and the Mactavator system and had some issues to work through as a result.

Week 11 | 14th - 19th December
The same as above for this week, the system was live on some of Thursday and all Friday, however, it was populating new agents, so we're excluding these results altogether. We also identified an issue that was suppressing selling agents this week, something we have only now fully resolved at the end of January 2021.

Week 12 | 21st - 26th December
Christmas came and with it some odd market movements and high volatility, it was also a short week. Despite this, we still ended up doing exceptionally well, especially because the system was still full of known bugs at this point.


Week 13 | 28th - 2nd January
Once again, we still had an unbalance between buying/selling agents causing sporadic results. Hopefully, we managed to resolve this in time for week 18, which was later than we hoped, but also came with some really exciting improvements.

DECEMBER 2020 UPDATE ARTICLE
The two main improvements for December was the addition of manual dataset selection and implementation of what we have dubbed the "Mactavator."
Simply put, the Mactavator ensures that we have a wide variety of agents that can perform in different market conditions, it does this with the following rules.

1) An agent must be the most activated at THAT datapoint (a higher activation = better understanding of the current market condition)
AND
2) It must have an equal or greater Total Return for that datapoint
AND
3) It must have a positive datapoint return (it must have won the trade it took at that datapoint)

There is likely a better implementation of this, however, it's only marginal so we're not worried about it for the time being. For more of what was accomplished in December, check out the full article here: https://www.mlforex.com/blog/update-...vator-is-alive

JANUARY 2021 UPDATE ARTICLE
In January we fixed a number of known and new critical bugs, these were the last clear roadblocks holding back the development of MLFX. We also rethought what our MVP (minimum viable product) would look like and started formulating a plan of action.

For more of what was accomplished in January, check out the full article here: https://www.mlforex.com/blog/update-...squashing-time

FEBRUARY 2021 UPDATE ARTICLE
In February we welcomed Christian, a front-end expert and machine learning P.hD to our team. He is a close friend of Amaury (MLFX CTO and co-founder) and valuable addition to our ranks. This sparked a pivot in product direction towards something more certain. We realized that our current trajectory had no definable endpoint, yes we had metrics in mind, but the amount of work required to get to those metrics was uncertain. We needed certainty in order to ensure the financial viability of MLFX long-term.

The new MLFX MVP objective
Previously our customers would spend hours scanning charts for setups that meet their trading criteria. A tiring process that could lead to errors in judgement because of fatigue or frustration.

Now MLFX scans the Forex markets for their trading criteria and notifies them when a setup is found, they then decide whether to take that trade in their trading platform of choice.

Our customers use MLFX because it saves them time, improves the consistency of their trade selection, and allows them to rate the effectiveness of their strategy against others including algorithmically altered strategies; thus enhancing their overall trading performance.

Our full February update can be found here: https://www.mlforex.com/blog/update-...our-team-grows

MARCH 2021 UPDATE ARTICLE
After discussing our objectives with Christian we decided on using Vue/Node of our web application technology stack, Trading Vue for our charting, and AG-Grid for our data grids. This will be a lethal combination that will position us in the top Forex application on the market. Yes it will take some time to get there, but we are playing the long game. We also incorporated our company and Amaury went full time on the project, two massive milestones.

Full article here: https://www.mlforex.com/blog/update-...nested-signals

APRIL 2021 UPDATE ARTICLE
Work on our web application began using Vuexy as a base theme, we migrated all our repos to Github in anticipation of more people joining the project and because that's where the open-source community lives. We also hired 2 additional developers to help, however, they both didn't work out, which was unfortunate but also a valuable learning experience.

Full article here: https://www.mlforex.com/blog/update-...-beautiful-vue

MAY 2021 UPDATE ARTICLE
This month our web-facing site mlforex.com reached what I would consider v1.0. It's a website I'm proud to take ownership of and will only continue to get better as time goes on. We're quickly closing in on the launch of our app minimum viable product ("MVP"), as an update, this is what our MVP consists of.

We have 3 types of strategies contained within MLFX:

1. Human Strategies - These are strategies found on the internet or requested by our users. We add these strategies manually and test/compare them so you can see how they perform. Unfortunately, due to legal issues, we might not be able to let you trade these strategies through our system, but that’s where the next step comes in...
2. Hybrid Strategies - We take Human Strategies and feed them to our agents, these agents use machine learning to test each human strategy in thousands of different combinations. They are then compared against each other and the original strategy so you can see how they performed. These strategies are unique to our system which means our users can follow and trade them.
3. Hermes Strategies - For Hermes Strategies, our algorithm has full control of selecting inputs and configurations for agents, their objective is to find the most profitable strategies, improving them iteratively with machine learning. Hermes Strategies will also be tradeable for our users, however, they are not part of the MVP and will come later.

As an MLFX user, you will use Hybrid and maybe Human strategies to assist your trading decisions. They monitor the markets and create “signals” when a trade setup meeting their criteria is found. You decide which signals you will trade using a trading platform of your choice.
We will quickly expand our functionality to include:

1. “Strategy Watchlists” allow users to follow strategies they like.
2. “Alerts” will alert users when new trade setups are found by strategies they follow.
3. “Trade Plotting” will plot trades for followed strategies on our live charts.
4. “Hermes Strategies” as described above

Full article here: https://www.mlforex.com/blog/update-...ing-humble-pie