• Blog
  • Durable models development principles

Durable models development principles

Share article

The cornerstone of trading models development is to lower the risk of losing money and the risk of future failures. Once we have built a model, it might work great on backtesting for a few years. Then it works fine on real money for a few years. But suddenly it stops working. 

If a model stops working immediately after the initial backtesting, EndoTech’s experts say that this model is overfitting, meaning that it only fits the backtesting period. And if the model stops working after the real money trading, such a downturn of results is called fading. After fading the questions come up: should we wait out such a period? Is there a chance that the system will start working again?

Are models black boxes?

These questions come from the fact that stereotypically trading models are treated as black boxes: the systems that can only be analyzed based on their inputs and outputs, without any knowledge of their internal work. For those quants who treat their models as black boxes,  overfitting and fading are inevitable as death sentences.

Meanwhile, the best analogy of modeling can be found in sports. Trading is a zero-sum game. There are winners and there are losers, and the models are trained on the past results. And then models compete. Sports results improve, new participants come out better, stronger, faster, and we need to continuously train our models to be able to compete. 

Let’s review how EndoTech’s approach helps create durable models. 

Train durable models with different market behavior

We distinguish eight market states. Each time market demands a different skillset from a model it marks a new market behavior. Again, in tennis, even elite slalom skiers will most likely lose to mediocre tennis players as there is no such thing as a universal athlete. It is also very hard to build a universal trading model. 

The school of modeling dictates to develop of the best “athletes” given the market season. Meaning to equip our models with the skillset required for a specific market pattern.  At each training session, we must define the goals to reach. Such as the longest jump as possible for bearish trends, the correct response to sidewise swings, and so on. That is why we need to find a way to identify the seasons to make models durable.

Apply filter to train durable models

Here comes a very handy concept of a filter. We trade only during the seasons that we find easy to identify and we have a durable model trained for it.

Our logic works as follows: as in calendar seasons, we can distinguish winter from spring by checking the average temperature and the natural biological clock. Same for markets, we can fairly easily distinguish a bearish market from a bullish one by using moving averages and volatility indicators. 

But what if we are too late or too early with our predictions? – There is another skill set that we have in our system: filter building or meteorology.

The main idea of trading is:


In the crypto market, there is a very consistent switch between seasons with rare changes in patterns. Bullish moves are followed by short-term bearish reversals, then sideways market, and then bullish trends again. After several bullish trend cycles, you can see a strong corrective bearish season. 

In the forex market, depending on the timeframe, some seasons are underrepresented such as smooth, bearish, or bullish trends. And even if we get into them, simple stop loss or time-based stop takes us out safe. 

Take into account new significant players

The only true turmoil for durable models is the new players that come to the market and change market dynamics. It might take time for model developers to understand it or, even more important, to accept it. 

When new significant players enter the market they can change the way the market moves and reacts on both macro and micro levels.

At EndoTech we have developed a few markers that are responsible for identifying market-style. It includes volume changes, rate of change, first, second, and third differential functions on each timeframe and element, as well as a statistical prediction for each price pattern. Once it starts to change, we downsize the portfolio allocation into models in this market. And again, our mantra: in case of lack of clarity, just do nothing. 

But sometimes you might decide that all your model needs are just to learn a new trick. 

Avoid overfitting of durable models

Imagine your elite skier had to perform during a foggy winter, so you decide to spend a year training him on this certain skill. But next winter is crispy, clear, and icy. So athletes fall short in these new conditions.

Overfitting means overspecialization.

On the other hand, you need to train your models the way elite athletes are trained by using advanced technology to help them understand small differences and small issues. For example, a tennis racquet for Roger Federer is completely different from another athlete’s one, because it is designed according to the details of his physique, his arm’s length, and other parameters. 

EndoTech successfully takes into account multiple factors to escape overfitting. We use a holistic approach and we try to reach multiple-goal functions. It means that we are building a model trained for every situation in the journey and we make it reliable for these markets. 

Making work technological aspects of the durable trading models optimization means taking care of multiple goal functions with a set of constraints. It is not an easy computational task. 


  • Black box approach to trading models development does not allow for building universal durable models that work in all market conditions
  • To achieve the risk reduction goals we need to build multiple models
  • We need filters to define the system and avoid overfitting
  • Trading models development is about training models all the time

Endotech Crypto AI


Share article