Are you sure that you have all the true details to judge his friend from a vauge post? Have you thought of the possibility that the OP is actually the friend and he reverses roles to fish what others think about the skills and the motives of the ATS developer he has hired? No, I guess you haven't thought of that possibility because you are seriously contrained thinking that far...I mean seriously....as you have already demonstrated...
You tried getting a holy grail ATS and gave up and then came up with some temporary strategies all in the 3 months? You haven't worked hard enough for that holy grail, imho, or do i misunderstand?
I basically decided after much research that there is no such thing as a holy grail. There are only opportunities that pop up and that can be exploited, but not indefinitely. Unless you have the scale or the power, that is. So I decided not to waste time on finding the holy grail but rather find something that works in the current time frame. My forward testing and live trading thus far has been quite profitable. It's just a matter of when to adjust the strategy.
What research is that, i'd be interested. There are traders out there who have been using the same technique for 10+ years and stayed profitable. Their technique is their holy grail. Now tell me one reason why you couldn't teach an ATS to do the exact same thing as that trader. A good ATS will be able to interpret all the information that a human can, and act upon it in the same way a human can. Whatever a person thinks is based on information and logic, and both of these can be replicated in code. Traders who do not act on logic, but on a whim, may simply be unable to describe their inherent logical processes, but it doesn't mean they don't operate on logic and rules. They also "adapt" their strategies over time, but a good trader will be able to explain to you how they adapted it, why and when. This can also be replicated in code. To me, the charts are a data set that is fairly simple to interpret. The trader and the ATS is pretty much getting 90% of the info from a set of bars with OHLC and volume data (news being the other 10%). It's such a simple data set that it warrants itself to be interpreted easily by "AI" within any ATS, which can thus easily copy human processes. There are also dedicated "news readers" applications that read the news and interpret them based on key words - this is the only bit of information that is a little more complex for a computer, and simpler for a human, but it has been overcome. So you see, in my view, if a trader can trade profitably for a long period of time, a computer can be taught to do the same, since a good coder can translate all thought processes into code. If you disagree, describe a thought process a trader may have based on information within a chart that cannot be replicated.
Here is my opinion on this matter. Markets are fractal in nature, and therefore are multiscale in nature. There are cycles in the minutes, hours, days, weeks, months, quarters, years, decades and ceturies timescales. Therefore, you might be able to find things that might work in those timescales, although the problem with the longer time scales is that you can never prove that you have a statistically significant result because you just don't have enough data. Markets are also non-stationary. You cannot expect the moments of the distributions to be fixed for all time. Humans are different from machines in 2 respects in regards to trading: (1) they are adaptive even though they may swear that they are using the same setups over 10 years, and (2) they can suddenly take into account things that were outside their consideration before. For (1), I suspect that over the years, good traders will slightly adjust their methods to maximize their gains. This might be an explicit change, or it might be a subconscious thing, where they are changing things gradually without even realizing it. For (2), you could imagine a case where one is trading happily along, and then some new commodities exchange rule comes in, and all of a sudden the trader observes a peculiar phenomenon that ties oil prices (not in a correlation sense, but in some conditional sense) with his success rates. Or it could be some other thing, real estate. Or the political climate. Bottom line is the market is an open, not closed system. So you cannot possibly code in every possible thing that could start significantly affecting the market. These are again my opinions, they are not facts.
OK well, I agree for the most part. I am probably assuming differently though, because I am working on daytrading ATS which executes around 10 trades per day on a single underlying. In my experience (coupled with a bunch of studies, tests, etc), the small-scale signals and rules that traders use on such a short time-frame have not changed for a very long time. True, there are some effects like 1) and 2) you mentioned that make the markets "open" in that way, but looking at intra-day data, not much changes because of them. Especially if you are focusing on a very general de-coupled and unspecific underlying such as SPY ETF, the ES futures, or even the USDEUR. On smaller time-frames like this, you also solve the first problem you mentioned: lack of data. With a 10-year backtest, with 10 trades per day average, that amounts to around 26000 trades to look at, and this is in my opinion enough data. The longer the time-frame (of having open positions), the less significant the results, so I probably made a mistake by not mentioning my assumption regarding the ATS frequency. Still, you are slightly wrong. A good system will incorporate all the data which is affecting the markets. Such systems are hard to come by, but they exist. That includes incorporating exchanges changing rules, margin requirements, or whatever consequences there may be due to any of those. Systems exist that model the changes such rules will probably have on other systems that trade a specific instrument or derivative. They are probably not something an average person will code within a few months, but it's doable. Often, they require human input though (such as how something psychologically affects the market participants) so it's a half-baked argument I'm presenting. Anyway, my main point was that if you are developing a system for daytrading which does something like 10 trades per day at least, the rules that exist on those intra-day 1minute charts hardly change. This is the first reason why an ATS is best developed for this timeframe. The other is that it gives you a huge amount of data to work with, thus increasing the probability of your hypothesis testing (backtesting) projecting forwards. This much I suppose we can agree on, and it is my advice that everyone should develop an ATS within such smaller timeframes.
Thanks for the thoughtful reply. I agree with you mostly as well. Regarding the high number of trades per day, I have found that if you don't account for commission and slippage, obviously there are plenty of profitable strategies that I can come up with. The PL curve just keeps climbing and climbing. But the minute I include realistic slippage and commission, pretty much all of them become unprofitable, with the exception of a few that seem to have these regimes of profitability -- profitable in a bull market, profitable in a bear market, profitable in a volatile market, etc. There is obviously a very small profit margin that requires exquisite fine-tuning without overfitting to get a consistently profitable system. And therein lies the problem. Regarding the statistics of number of trades, I don't believe you can simply count the sample size and say that is a large enough number. That is because during a single market regime, those samples will be correlated in some way (e.g., during a bull market, the price action will behave slightly differently from a bear market). So this is why I think there are regimes during which algorithms work, and others where they don't. Of course the logical conclusion from this is then why not fit separate algorithms to separate market conditions and choose which one to use based on the market conditions of the last month. The problem is this: we can never anticipate all possible market conditions. The simple classification of bull, bear, volatile, non-volatile is likely not enough to cover all cases. For example, this last 3 months, I noticed that there is huge dependence on morning news, much larger than say prior to August. Of course, you could lump this in with high volatility, but I also noticed that March and June of this year were days that were particularly insensitive to morning news. Of course the ATR during that period appears depressed as well. Regarding all-knowing systems that code in every possible scenario, well I just have to point to the flash crash, or the 2008 melt down to show that once in 10, 50 or 100 year cycles can happen and ruin everything. Even the most complete systems in the financial industry appear to have missed them. And as far as those who made money off these rare events, it is difficult to prove or disprove whether they got what they planned for or they got lucky. Now, bare in mind, I have never considered say examining the level II order book data and playing games like that. I am strictly looking at price action. So I cannot claim to know whether it is possible to come up with rules based on the order book and be consistently profitable.
Well hmm, I think a good logical trader will be able to describe to you how he changes his daytrading rules based on conditions. I am confident these can be implemented so that a single system slowly blends between different rules for different market conditions. Now this is my conviction, based on information i consider credible, and i haven't proven it. So far, i've managed to create systems that adjust and blend values (parameters, input), but i cannot comment on profitability. We'll see in the real world what happens, but this is an interesting discussion. Money management makes up 80% of a strategy (if not more) and the strategies main outputs should be entry/exit and MM adjustments - where MM adjustments are probably more important and easily ignored. I believe there are daytrading strategies that can adjust to varying conditions and not screw up the MM on "odd days". When a backtest is run and we pick the worst days, it's easy to see what combination of rules didn't work and why. For me, one important aspect is developing your own strategy development platform - from scratch - and outputting sets of values into indicator areas under the chart to monitor exactly what your ATS is doing and why. I think with this approach it's much easier to detect hiccups, adjust based on market conditions, and then test and see if other market conditions have started underperforming as a result of the new modification. This may sometimes lead to an infinite feedback loop of adjustments and subtle fitting, but this is where creativity and inventiveness comes into play. It is somewhat irresponsible to NOT include commission and slippage in simulations from the start. I've included them since i built my initial framework, and the slippage is pretty ruthless in my default setting - also depends on the speed and volatility of the given simulated ticks. I think you should test all your systems with the same assumptions and a robust slippage model too. This is pretty crucial. Regarding Level2, foreget about it. It is filled with ghost quotes and isn't much use unless you have exchange-level data access which is expensive to say the least. In the end, I have to say I cannot confirm my beliefs until i have a fully working and tested system in real time. So maybe in the end, i'll switch back to the idea of running different systems for different market conditions, but as of today, i think i am on a good path to create one that works consistently. It's also a bit of an academic challenge. Regarding your original topic, i'll have to say you sound like you are able and should be fully independent in developing your ATS, so dump that guy and stop him wasting your time. Good luck.