I'm an engineer doing academic research for my master thesis in the area of quantitative finance, basically the purpose is to study the possibility to create an intraday-trading algorithm.

I've tried a regression algorithm (SVR) to predict future prices without success and currently I'm using Boosted Decision Trees (BDT) for a classification problem (Long/Short/Out) where I would take advantage of a daily trend (place an order at the beginning of the day, sell at the end of the day), for example:

• Long when BDT predicts Close > Open*(1+a)
• Short when BDT predicts Close < Open*(1-a), where 'a' is an error gap;

I'm using technical indicators: RSI, MFI, SMA, EMA, MACD, ATR, Bollinger, and linear combinations of them. I think I'm doing everything right, all the indicator values are normalized with the Open price, I'm using cross validated metrics and a grid search to look at different combination of parameters for both the technical indicators and Boosted Decision Trees.

But until now it seems that, at least at the intraday level, the market is just a stochastic process! I have also read a few articles available online and found a couple of errors on them (using lagged data or indicator values, for example), which leads me to believe that most of the literature is just junk, I mean it's impossible if everyone that tries to create a trading algorithm that generates positive returns is successful, right?! I won't even talk about the ridiculous articles using technical analysis with the supports and resistances, I generated a Geometric Brownian Motion plot and I also see those hypothetical supports and resistances, but there is no rational behind them, just people's imagination of things that don't exist.

What is your opinion on this subject? Do you have knowledge of a successful intraday trading algorithm? With what kind of average daily returns (0.01%,0.1%,1%)?

• If I had that knowledge I would be making money and not posting the strategy here. In any case, I am an academic, I believe that markets are efficient and that although there exist frictions that allow to generate alpha those frictions for trading at high frequency are not exploitable by any individual investor. In any case I think your question is offtopic here. Apr 1, 2016 at 11:53
• I'm not expecting people to post their strategies here, I'm just curious If the research I'm doing is worthless or not! I hear that there are hedge funds using these types of algorithms but I've never seen/read anything in detail... Apr 1, 2016 at 11:57
• Yes, they use. But at such high-frequency only computers can trade in a profitable way. So for you is pretty much worthless. You are better off in finding a strategy for the long-run. Read about Navinder Singh Sarao and also why you will never beat the trading algorithms of wall street: telegraph.co.uk/finance/newsbysector/banksandfinance/10736960/…' Apr 1, 2016 at 12:00
• I think your general impression is correct: much that is published or marketed on this subject is trash. "The people who know aren't talking and the people who talk don't know". But it is quite difficult even for knowledgeable people using real money to be successful in this area. Markets are reasonable efficient, what works in sample is not guaranteed to work out of sample, etc. Apr 1, 2016 at 14:44

Such a complex question...

Geometric Brownian Motion (GBM) will not typically work to aid one finding strategies based on technicals, as the pursuit of the technical trader is to find market deviations from a random walk.

However, some strategies, for example a "take profit/stop loss" strategy can work, (or at a minimum one can change the risk/reward profile) using GBM on assumptions of limited slippage (where stop loss is not effective due to jumps in price). This is due to the non-linearity of likelihoods vs risks/rewards.

The typical problem of empirical findings vs real trading involve overfitting of data/models when deciding upon a strategy. And if a strategy is successful, how does one know it's because it's a good strategy or anecdotally worked subsequently?

If one gave 10000 monkeys buy/sell buttons their results might approximate a normal distribution. If one took the top performing monkeys and gave them buy/sell buttons, some of those would perform better than others. Can one say anything about the top of the top performing monkeys?

Similarly, if one tests enough strategies on a set of test data, some will work admirably. And if data is split into 2 sets? One to test strategies, and a strategy chosen, a second separate set of data for confirmation? Well, done enough, again some will perform better than others. These are examples of 'data snooping'. It typically proliferates - from articles online, even into literature - books and academic papers. So data rigorous data splicing is required, but this can present its own problems, including correlations of some attributes between data sets.

Are successful intraday strategies possible to find?
My belief is yes - based upon my experience working in investment banks, and my own research. Highly dependent upon market, and external factors are usually required to refine strategies to a point where they are profitable. There are a plethora of books and papers suggesting things such as momentum, sentiment, auto-correlation, mean-reversion, trending, and oh so many technicals. How can they work? Because in reality the market is not always random - there are periods of non-randomness. Sometimes these are tiny squeaks obscured by a concerto of noise (particularly short-term). But, if you can find these pockets of non-randomness, you can find successful strategies.

If you don't believe me, then listen to the man from Renaissance, one of the most successful Hedge funds of all time: Jim Simons Interview link

Note: I'm not suggesting they use "technicals" in the typical use of the word. They might, or not. I've no idea what they use, but I believe they use models based on more than just fundamentals.

• Thanks a lot for the answer and for the link, I really enjoyed watching the talk! This was what I was looking for. Based on the talk I suppose they use a lot more than just technicals and fundamentals, they were talking about the weather and other date, basically they try to incorporate all the information that may influence human errors or judgement decisions. Now it starts to make sense ;) Apr 1, 2016 at 23:24
• My pleasure, I'm pleased you enjoyed it, and I for one would be fascinated to see your thesis! Good luck with it, and may you find some non-randomness ;) Apr 2, 2016 at 2:16

Here's my favorite example of an intraday strategy on S&P500 futures that at least used to work:

Intraday Share Price Volatility and Leveraged ETF Rebalancing

I pull it out whenever people start talking about market efficiency. The strategy is very simple: if S&P500 futures are up or down more than 2% on the day with two hours left until close, follow that direction until close. At first sight it looks like a random non-sense strategy, but if you read the paper you can clearly see that there are some very predictable Leveraged ETF rebalancing flows that move the market.

I would say that it's definitely possible to find a profitable intraday trading strategy, but you have to be able to answer this question: "who is losing money on the other end?".

You can see another entertaining example if you search for news on "Good Harbor Financial".

• yeah, it's been quite a while since it has stopped working :) Apr 1, 2016 at 17:07
• Thanks for the answer, I'll read the article it seems very interesting! Sorry I can't up vote yet... Apr 1, 2016 at 23:30

I have been through your confusion myself for the last five years. Until recently, my account started to get some consistent performance.

1. First, I started with Technicals, Spent \$ on a automated trading platform. From there I created common strategies. The results is not promising. The strategy doesn't consists parameters and if one strategy works on one specific product, it probably won't work on the other product. The movement of the price seems so random. And my expectation for the strategy performance is high. (Nearly did I know I was so close.) I concluded all the failures due to simplicity of technicals.

2. I then moved onto next stage which tries to use complex model to predict price movement. The complex models includes using various machine learning techniques. That doesn't work well either. The prediction accuracy is not as high as I expected.

3. Then I thought I might need some quantitive finance knowledge to better understand the financial markets. I then studies stochastic calculus and various quantitive financial techniques. In the end, what quantitive finance tells me is that I should get the same return as the fixed-income. And the time series predication model doesn't perform any better than my machine learning techniques.

4. At this point, I considered giving up. Because each of above tasks takes tremendous effort and time. The most hurting part is that, after all these efforts, there is no rewards.

5. At some rare situation, I met with "this investment group" of people, they shed some lights that brings me to today's state. Market is mostly random but there are some non-randomness in it. And this non-randomness is your chance of making money. But it's buried under high randomness. The expectation of making money should not be high (you'll never get an ATM), but by careful control of risks and leverage, it's possible to make enough money that others can only be jealous about. Slippage and Transaction cost is your enemy. On average, you can only get so little after slippage and transactions cost. But for financial assets, there are leverage that the little can be huge. Leverage is a double blade sword. Risk is also increased under leverage. But the good thing about risk is that there are some established practice in both academia and industry that can help alleviate risk (but not completely remove).

Finding a strategy is only part of your job. Finding a "good" strategy is hard, but finding a "working" strategy is not so difficult. The other part of your job is to reduce the risk using various methods. Then increase leverage to get more rewards even though you have a "mediocre" strategy.

I hope this can shed some lights to people who are still struggling alone.

• Thank you very much for your answer, it seems like I'm going through the "normal" path to finding a profitable trading strategy/model since I'm almost 2 years into this. As you said sometimes It's a struggle and seems that the market is just pure randomness but I've recently developed a very promising asset allocation (and stock picking) model that beats the benchmark index every year so I'm starting to see the light! I'm still struggling with the trading model itself, I think that if I could join both models I'd have a winner! Dec 19, 2016 at 10:14

I know this question is quite old, but I just wanted to mention one small problem I noticed in the past about using ATR as a feature input.

While other technical indicators like SMA or EMA can be exactly replicated for a given timeframe (as long the timeframe is longer than the timeperiod used for the corresponding indicator) the ATR (Average True Range) and NATR are based on all previous values. To be more precise its based on one previous value, but in a recursive way.

Therefore you would have to either recalculate the ATR from the beginning of your timeframe each time (the same timeframe you have trained your model with) or save the ATR indicator together with your data and calculate all new ATR values based on the past ones you saved.

Depending on how you build your SVR model that could be one reason why it failed. The bigger risk with this problem is it couldn't appear while modelling, because all indicators are calculated correctly in the training process, but could then appear later while deploying the model (like it did in my case).

I'm writing this because I couldn't find any information on this problem in the past.