How to find a good trading model in foreign exchange trading

I think that for a relatively mature trader, the procedural entry point is often to objectify their existing system first. For a novice or a trader whose existing system is difficult to quantify, the procedural entry point is to start by imitating some simple but classic models.

Then comes the test. Test for what?

It should be possible to see the interface related to parameter optimization on every common platform. At least in China, "parameter optimization" - maybe this is a widely concerned and controversial issue. Many newcomers will be confused about this at first, because many programmatic people are against parameter optimization. In fact, what I first heard was the drawback of overfitting, which is easy to accept - the more fully fitted to a certain period, the more likely it is to deviate from another period, because we always fit only certain features, that is, our fitting is destined to be incomplete, so future deviations are inevitable. But then I heard "against parameter optimization". I feel that this is a bit of a narrow-minded. Is it against this button function on the software? So many software companies want to set up such a function, there is always a value. Is it against trying to optimize the idea? But if you don't optimize, how to set the parameters? "Rely on common sense"? How does common sense come from? How much common sense do newcomers have?

For example, let me give an example of the properties of N-ethylpyridine bromide. Not to mention the parameters of color and specific gravity, you don't even know whether it is a solid or a liquid at room temperature. If you are not engaged in room temperature ionic liquid research, how could you possibly know? Programmed trading is not something that everyone has experienced and experienced since childhood. It is a brand new thing for newcomers, and there are still many things that the elderly do not understand. So I think relying on common sense to determine parameters completely violates the cognitive laws of the human brain. Newcomers have empty heads and can't shoot much common sense. I feel that this may not be beneficial to programmed trading, because even if overfitting to one period can produce excessive deviations in another period, if the existence of excessive deviations is recognized in a certain period, it will definitely be exchanged for an excellent result of relative fit in the future? Is there such inevitability? I am afraid that the parameters that are randomly determined will not fit this period, and that period will not fit either. Before excellent performance, the principal has been greatly reduced, and how can traders have long-term beliefs? It is the last word to pay attention to the details of the specific performance of a certain set of parameters and understand its possible implications for the future.

In my opinion, parameters are often the eigenvalues of the system, and are mainly eigenvalues that measure market volatility. The main purpose of parameter testing is to try to find meaningful eigenvalues, and meaningful eigenvalues are measured by the dependency stability, historical stability, and universality across cycles and varieties. Let me give an example to illustrate these few abstract words. First, the theoretical premise is that market volatility is the result of a combination of waves of various frequencies. Those who have some basic physics should know that the description of the synthetic wave can be solved by decomposing the wave into the fundamental wave and all levels of harmonics, and by describing the frequency, amplitude, and phase of each sine wave separately; or we know the frequency, amplitude, and phase of the fundamental wave and harmonic, and we can reproduce the original synthetic wave. These frequency, amplitude, and phase representations are reflected as parameters in programmed transactions. If you find this parameter, of course, it has high stability and universality. The origin of these decomposed waves is very diverse and complex, and it can even be said to be inexhaustible (because of its nonlinear market nature), but we can still clearly observe the harmonics corresponding to some linear characteristics.

For example, for industrial products, demand is usually growing steadily, and the diversity of demand is characterized by the changeability of the demand side. But the supply side is relatively simple. Under the economic cycle or policy cycle, there are often investment peaks, and the investment peak will turn into the production peak after a certain period of time, causing a wave of decline. From raw materials such as crude oil minerals to intermediate industrial products such as PTA, it is very obvious, and due to the domestic government's macro-control, this investment imbalance has actually been exacerbated. For many varieties, there are naturally off-peak seasons. For example, clothing has two seasons: spring, summer and autumn and winter. After a fashion show every year, there is large-scale mass production. After mass production, there is an increase in cloth procurement. The increase in cloth procurement continues to be transmitted upstream, corresponding to the increase in yarn and cotton procurement.

Another direction is the increase in the procurement of polyester staple fiber, polyester filament, elastane, nitrile, viscose fiber, etc. Further up is the purchase of PTA and MEG by polyester factories, and further up, the purchase of glacial acetic acid and PX by PTA factories, and the purchase of naphtha by aromatic hydrocarbon factories... These procurement times cannot be too different (inventory cost constraints), nor can they be completely divided into daily (logistics economy of scale considerations), so it is destined to be reflected in waves of procurement peaks. There are procurement cycles in all industries, but the strength is different. These procurement cycles are determined by logistics efficiency, information systems, and management methods. At present, it is the era of electronic networking, usually 2 months; but before networking (ten years ago), there were also many local networks that could not be covered by roads and railways, and the road conditions were also poor. At that time, the warehousing and procurement cycle must have been long, 3-6 months; further back, the era of horse-drawn carriages and sailboats two hundred years ago must have been longer. A cross-ocean cross-border trade (such as transporting spices from the Nanyang Islands to Rotterdam) took more than a year, and the amount of procurement could also be used for two or three years.

In short, this cycle is determined by the level of economic development, which is a wave. This wave is a medium-term wave, and the non-linear characteristics are relatively weak, which means that it is easier to detect and measure. It changes very slowly. You can't see it change significantly in five years, and once it changes, it is difficult to change back. But another wave that often superimposes on it is the fluctuation caused by psychological expectations. Buyers always want to buy at a bargain, and sellers always want to stock up and wait for the price. So buyers tend to move the phase of this wave forward, while sellers want to turn this wave into a square wave as much as possible - that is, the longer the high time, the better. The supply and demand relationship and macro environment vary from year to year, and all parties are accustomed to the rhythm of the previous two years as a reference, so you will find that although there is a range of price highs every year, it must be different (spore theory). For example, if the previous year's copper high point appeared in April, and there was only one obvious high point. The buyer will think that if I buy at the end of March, I can buy cheaper than the competition. As soon as the seller sees the increase in inquiries at the end of March, he immediately raises the price at the end of March, resulting in a price high in March. Due to early purchases, April may be out of stock and fall.

Of course, if the variety is in short supply that year, then purchasing in advance will also lead to early use up, and at the same time, coupled with the demand for short-term in the early stage, so after the fall in April, there may be a wave of concentrated purchasing actions to replenish inventory in May. In this way, there may be two highs in the spring. This is the rise and fall of psychological expectations, which is caused by the superposition of the original procurement cycle. Psychological expectations also have a rhythm, but they are volatile and self-feedback, so they have strong non-linear characteristics. But this is not the final market rhythm you see, because there are many big and small factors that cause fluctuations superimposed on it. Finally, there is a background noise that exists all the time...

The above is just an example to show that it is possible for us to find several stable and universal parameter values, and the values found are realistic, but because the reality has been repeatedly distorted, it is impossible to guess a meaningful parameter value directly by observing reality. But parameter testing can help us find their values, and through stability and generality testing, their values can be finally determined. It is roughly the same order of magnitude as the value we guessed to be realistic. Parametric testing can also tell us the degree of generality and stability of these meaningful parameters. For example, some are universal across species, but there is no proportional relationship across time periods.

Another possible historical stability is acceptable, but it can neither cross varieties nor cross cycles. Even some historical stability is not possible, but its historical response has a single linkage law with a certain correlation characteristic. Parametric testing also allows us to know which parameters have not found a definite practical significance, but are dependent on another dynamic variable (often the embodiment of breed specificity, seasonal specificity, etc.). This understanding allows us to take specific adjustment directions for certain correlation parameters when dealing with different varieties, different time periods, and different market styles, and this is the basis of adaptive parameter methods and dynamic parameter table methods; even if you adhere to the semi-automatic road, this understanding also provides a solid practical basis for subjective screening at the implementation level. Instead of the embarrassment of thinking about changing the model, but not knowing where to start, or changing it casually. What's really elusive is those parameters with poor stability, poor universality, and no single direction of relevance. But it doesn't matter. Just give a discount.

We can qualitatively judge whether the model is a positive return model. If it can get an annualized rate of return of 30% and a maximum drawdown of 20%, but there are two internal parameters that cannot find practical meaning and change laws, we think it may not be profitable in the future, but after all, it has made money, so stick to it, it is expected to become a breakeven model. If there is a way to turn it into an intraday short-term model, the risk will be controllable, and the speculation fee will be more appropriate. Even if you don't value this piece, then at least treat it as a combination object to smooth the fluctuations in the performance of other strategies. In fact, maybe one day you will find the practical meaning or law of the parameters, and then you can use it more boldly. Keep it in your strategy prototype library first. What does a programmatic trader think about every day? Largely studying this.

What else is the time thinking about? Logic testing. In fact, logic testing is more important than parameter testing, and the improvement effect is better. If parameter testing tests the level of computer hardware, then logic testing tests the strategic designer's understanding and creativity of the market. But the two are inseparable. In fact, some parameters, when taken to an extreme value, are equivalent to blocking a certain logic.

The entry point of the logic test is the microstructural analysis of the market and the microstructural analysis of the logic itself. The keyword of the former is to capture, and the keyword of the latter is to split.

Capture can be divided into active and passive, and it is difficult to achieve active capture on existing common platforms and my mathematical knowledge. Moreover, our personal computers with existing programmatic platforms cannot run too complicated algorithms and logic, otherwise the calculation will become a computer shock. But an example can make everyone understand the meaning of active capture. Human-machine voice dialogue is actually to identify features through active capture through the acoustic laws of each syllable to complete vocabulary restoration. It is most appropriate to let a pattern recognition expert engaged in waveform analysis do this. Passive capture is a stupid method, but it does not have to take many detours, but it will often make the tester feel incredible, and the uncertainty of the result is sometimes frustrating and sometimes surprising. Passive capture actually uses logic and parameters to measure a certain feature. The logic and parameters are basically limited, and the microstructure of the market shape is also determined. We can only find a stable value with high mathematical expectations to show that we have successfully measured and obtained a realistic feature. It is like catching fish in the water with your eyes closed.

I have ready-made examples. Some varieties often have multiple tops and bottoms, but they are irregular. To identify these complex and irregular K-line combinations, it is really not possible to achieve through simple logic - this was initially tried to achieve directly, and I gave up thinking about it. But later, when studying how a certain parameter controls the behavior of the model, after splitting and optimizing, I found that a simple logic can not fully capture this morphological feature, and the introduction of this parameter was based on a completely unrelated idea. At this level, I fully appreciate the chaos of the world: if you try to achieve precise results with precise settings, you will often fail. However, the vague setting should still be different from blind trial and error. The understanding of the form and logic itself is still a necessary prerequisite. The tests provided by various software platforms and outside the platform (such as through EXCEL or MATLAB) are the means to obtain this understanding. "Optimization" tests are nothing more than mathematical expectation tests that set a certain orientation (rate of return, maximum retracement, etc.).

Splitting of logic. This is a principle. When designing, each link should be designed as much as possible to be measurable individually. Shield the association between other parameters and logic, and only use this association after determining the practical significance. For the permutations and combinations we will mention later and the cross-logical combinations we will mention later.

The direction of logic testing: there are two types: leak detection and filling, and permutation and combination.

Missing and filling gaps often focus on dealing with extreme market conditions, such as stops, continuous stops, huge overnight gaps, etc. Especially minute line trading strategies. Unless it is foreign exchange trading, the continuity of the market is always destroyed by cycles, which brings considerable tests to the effectiveness of short-cycle strategy testing. Allocation and combination is the repeated trial and error of the logic that one can understand and cannot understand, which is a bottomless thing to do. There are many plausible things, and there are many plausible things. My experience further taught me that the design of the head is not feasible. There is no shortcut to finding a strategy that can be used in practice, just keep trying. I thought that programming could liberate the human brain from the ocean of information, but I didn't expect that programming is also a physical task - coming out of the ocean of information in fundamental research and plunging into the logical ocean of programming research.