On 1 August 2020 we’ll release a completely revised version of the D-Score.
In this post we explain:
- The origins of the D-Score
- What we’ve learned over time
- The new approach of the D-Score
- Effects on other features
- Darwinex’s use of the D-Score
TL;DR – The D-Score calculation is now much simpler, using only Quote data with zero dependence on Investment Attributes / Scores. Darwinex to continue using D-Score for risking proprietary company capital.
1. The origins of the D-Score
The D-Score was born in 2013, long before the launch of our talent and capital Exchange. At that time, our goal was to automate the identification of the best traders in the world. The intention was to give them visibility so that they could attract investor capital on the future Exchange.
At that time -even now- most signal services listed strategies using simplistic criteria. Really good traders just didn’t stand out. One had to have trading knowledge to find the proverbial needle in the haystack, a source of great frustration for both good traders and investors.
If we found that algorithm, we could empower non-trading clients to be successful. Talent, on the other hand, would perceive us as the only technology provider that genuinely cared about them being successful. The latter is something that, in our humble opinion, we’ve achieved.
So we started with this goal in mind. Along the way, we discovered that reliability of returns heavily depends on risk management. In order to normalize results, it was necessary to consider risk at the level of each position taken on the market by a strategy, and not simply the risk of the strategy as a whole.
A strategy opens 9 trades with 1:1 leverage that loses 10 pips on each of them. Then opens another trade with 100:1 leverage and gains 10 pips. The signal generates positive returns. But is this a good signal? Also, what leverage should we allow for the last trade to consider it representative of the previous sample? 5:1? 10:1?
The clearest and most common case of this type of trading is the martingale, but there are many variants.
This led us to create an algorithm capable of eliminating the possibility of statistical illusions. We also discovered that the best way to market the signals is via a product that enabled them to be comparable on an apples-to-apples basis. The DARWIN.
A DARWIN is the transformation of a signal into another where the individual leverage per trade or position doesn’t generate false illusions of profitability.
Alongside creating the DARWIN, we continued working on the talent-identification algorithm.
Besides measuring signal quality, the algorithm had to meet the following criteria.
- Provide investors enough information to enable confident decision-making, while protecting the intellectual property of the signal provider.
- Be dissectable into separate units. This is how attributes were born! We wanted to offer traders transparency and ideas for improvement.
- Be independent of the risk level.
- Be valid for any type of asset, as long as the risk was only a function of its volatility. This made us exclude options from eligible asset classes.
- Be valid for any type of trading operation.
- Be impervious to deceptive practices.
It’s important here to emphasize that the D-Score was born much earlier than the DARWIN asset class itself. More than 2 years earlier in fact. In spite of this, we continued working on it once the asset class was born as we conceived it as independent from the risk level.
2. What we’ve learned over time
What if we started developing the D-Score now?
“Now” meaning after several iterations of the asset class and more than 10,000 created DARWINs?
We’d most likely develop a different D-Score.
The complexity of the D-Score we’re leaving behind is a consequence of its evolution.
Now that we have the data, now that we’ve performed numerous analyses, we have come to the following conclusion:
Using only DARWIN quote data we can distill a more predictive D-Score than the current one. There is no need to use the investible attributes.
This is so because all risk normalization is already factored into the DARWIN quote. Hence the quote is a noise-free signal.
The D-Score we’re leaving behind was an important step in the classification of financial assets as it allowed us to compare strategies of different risk levels. But in the presence of noise-free signal, it is better to use only that to determine the quality of a DARWIN.
The constraints we imposed on ourselves in creating it, especially the need to calculate it from independent attributes, has prevented us from iterating the D-Score and its attributes as much as we would have liked.
The diversity of factors that influenced it (attributes, weights, history, assets) had brought it to a point of no evolution.
We all knew that we needed to improve, but we didn’t know how.
The only possible way was to simplify its calculation and make it independent of the attributes.
These years have also shown us the value of investible attributes. While they don’t have predictive capacity on their own, they do help to understand the type of the trading strategy.
Capacity, timing, loss aversion, market correlation, etc. are indeed metrics of great value, and we’ll continue to maintain and improve them over time.
Now that they don’t influence the D-Score in any way, we’ll be able to improve them without the fear of knock-on effects on the D-Score, as was the case in the past. Therefore, we expect a lot of value creation on this side.
The new approach is to have:
1. a value that, for Darwinex as an investor, determines the potential return of a DARWIN: this is the D-Score and,
2. attributes that describe such performance.
Investors are of course free to use them (or not) as they see fit.
There will be DARWINs with a D-Score of 80 and Mc=2.
Presented this way, the D-Score and attributes provide more valuable information than they do at present, affording Darwinex a more neutral position than before.
3. The new approach of the D-Score
Our research has led us to the following conclusions.
- A DARWIN’s Quote (price) can be used to determine its ability to generate future returns.
- Investment attributes provide information on how strategies achieve their returns. This is useful, e.g. for describing the behaviour of a strategy without the need for traders to describe it themselves.
- The best strategy to invest in DARWIN assets is to invest in those that in the medium term (2-3 years onwards) have been able to generate returns, and currently have a positive momentum.
- DARWINs that stop performing change the ‘‘shape’’ of their upward curve. In trading jargon, the breaking of a trendline is usually an indicator of strategy exhaustion.
Points 3 and 4 are what we have attempted to summarize in a single metric, the D-Score, and the results are promising.
Going forward, improving the D-Score will be easier. We’ll announce changes as a change of version.
Being considerably less computationally expensive, the D-Score will now be updated once every hour.
Based on data from all active DARWINs taken on July 15th (activity in the 21 trading days prior to this date), the mean D-Score is 39.70 according to the new version versus 35.51 according to the old version. The DARWIN with the highest D-Score has a value of 88.80 (new) vs. 85.80 (old). 62.07% of the DARWINs improve their D-Score. In the group of the 50 with the highest AuM 66% improve it and in the group of the 50 with the most investors 62% improve it.
On a distribution curve, the comparison remains as follows:
4. Effects on other features
4.1. Investable attributes
As mentioned earlier, we’ll continue to maintain them but they won’t have any effect on the D-Score.
We’ll soon discontinue the Performance attribute as it’s effectively redundant in the presence of the new D-Score.
Our plan is to simplify the calculation of the attributes, improve calculation frequency, and add additional attributes over time.
To simplify their use for non-trading clients, we plan to create a short automated description of the strategy from all its attributes.
4.2. DarwinIA Capital Allocation
We’ll continue to calculate rankings as before, but using the new D-Score. It will still be possible to land an allocation with a track record < 9 months, but never in the upper 50th percentile. Getting into the top 3 will be much harder, which we consider fair.
Starting August 2020 we’ll increase the max. monthly allocation from the current € 6M to € 7.5M.
In future editions, we’ll likely restrict participation if the Rs value is low. Rs is the only attribute providers should take care of to improve the quality of the resulting DARWIN.
4.3. Accounts without DARWINs
As the DARWIN quote is not available for accounts without associated DARWINs, D-Score calculation will be simpler. There may however be discrepancies between the D-Score value of an account without a DARWIN and of the value that same account has with DARWIN. Its calculation frequency will also be lower than for accounts with associated DARWINs.
4.3. Predefined filters
We’ll change some predefined filter criteria to adapt them to the new reality of D-Score.
4.4. D-Score rebates
We’ll raise the D-Score required to get a 20% discount on brokerage fees to 55 and maintain the limit of 60 to get a 40% discount.
5. Darwinex’s use of D-score
The new Darwinex Score will be -and we’ll communicate it as such- the ultimate metric we use at Darwinex to risk our own capital:
- via the DarwinIA Capital Allocation (a programme to train providers to operate a strategy that has third party AuM), and
- via the proprietary portfolio of DARWINs in which we invest our own capital.
So, the D-Score will be relevant for those providers who pursue Darwinex AuM.
The D-Score is not a recommendation to invest in DARWINs nor a guarantee of future profitability.
P.S. We’ve published a tool to consult approximate new D-Score values per DARWIN so that providers know what to expect beginning . You’re welcome to post your comments and feedback on the relevant forum thread.