Re. 20 samples being enough to achieve statistical significance. That is nearly true (a bit small for robustness in most circumstances) but the key is that it has to be REPRESENTATIVE data. If you have 60 days and 10,000,000 intraday data points, if it is a particularly quiet or particularly volatile period, it will not be representative of the true universe of possible distributions of prices that are likely to occur in real life. I teach statistics, and I often use this analogy: If I am doing a study on the height of average college students, and I decide that it would be convenient to take my sample at the dorm nearest my office. I walk into the cafeteria and survey 70 people, only to find that the average college male is about 6'10" and female is about 6'1". Of course, I have managed to hit the men's and women's basketball teams' training tables. I then do the same study with only half as many subjects, but selected at random over a 10 day period as they get out of their cars in the parking lot, and find that men average 5'10" and women 5'4" tall. Price distributions in the markets are kurtotic, with "unlikely" events occuring more often than they "should" in lognormal distributions. Thus, the quality of your data sample, and how closely it represents the true range of what typically occurs in the real world is often far more important than simply the quantity. In fact, a large sample can achieve statistical significance while detecting effect sizes that are too small to be meaningful in the real world, and sometimes a smaller sample can actually serve your purposes better, by revealing only the more dramatic effects and producing less "noise". Hope that isn't totally confusing....... Jessie
...I did not take your kind comments as critical. If I appeared to do so it is because of my pedantic writing style. I very much value your comments because they are coming from a perspective so different from my own. FWIW, apparently also I miscommunicated about my trading style. Once I have backtested, I mostly trade the system rigidly mechanically. I always take the entries and the stops, but I will sometimes use that cursedly uncodeable intuition to take profits earlier than the system call for. May I ask, is it your goal to develop systems which are robust across all time frames and instruments? If it's not proprietary, I would like to hear what the general features of your approach are. You said I don't need all the tools for proper backtesting. What in your opinion is proper? Thanks. - Mike
..aha! A statistician! Who trades! I failed that in college because I always tried to make the connection between statistics and probability. I mistakenly thought there was some causal connection. Could you comment on this method, please? It seems so patently obvious to me that it's probably wrong. I trade only 9:30 to 10:30 EST (after which the lamentable necessity to go to work intervenes). Therefore from the 60 trading days of history available I select only that time period to test methods that I want to work during that period. Sixty minutes times sixty days equals 3600 data points. The question I ask therefore is "What is the average height of the people who eat breakfast in the cafeteria". Do you have a statistically significant statistician's opinion on the statistical significance of such a test? You speak of price statistics. I think what we want is not the distribution of prices but the distribution of price sequences. I try to filter out the ho-um and jump on the profitable events (is that what they call momo trading?). And please, please send me the kurtoses so I can let them run and stop out the lognormals. If you have any market statistical insights to share that aren't proprietary I would love to hear them. Best regards. - Mike
As I see it, (and for God's sake, take everything I say with a grain of salt and realize that no two statisticians will ever agree on anything) the problem with the time frame you propose isn't necessarily the time-of-day issue, that might be fine, it is the brief 60 days of history. One 60 day period (even of just the brunch hour) might not (probably doesn't) include a 3-sigma event, and another might, perhaps more than one. The first will significantly underestimate risk and greatly inflate returns, the second will significantly overestimate risk and underestimate returns. But somewhere between none and several is the actual (probable) likelihood, and your model had better account for it or it will ruin you. Your sample has to be as representative as possible of all of the possible price distributions you will encounter. It may only be 20 data points, or it may be 20,000, each has advantages and disadvantages (e.g. statistical power vs. effect size), but they need to be both randomly selected and representative of all (or as many as possible) of the market scenarios you will encounter for best efficacy. It is unlikely that you will ever have a "representative" market period that only encompasses 60 days. As for the kurtosis issue, that is the holy grail of risk management for a lot of traders and their modeling. I can't really help you much with specifics, as there is no general agreement on how much correction to allow, and in addition, it is clearly not a constant across markets, nor across time frames. I can tell you that more diverse indices are less prone to extreme events than specific stocks and narrowly traded markets, but then that becomes reflected in the shape of their distributions, so it is somewhat accounted for there already (although usually not completely). Anyhow, if you figure out a short answer to that one, let me know! (but don't tell anyone else....) Good Trading, Jessie
...thank you very much for that definitve reply. I believe I understood everthing you said, and it sounds reasonable. When E-Signal provides more data I will test different durations on my favorite system to try to see which is more important, the rare event or up-to-date volatility measures. I suppose that the critical issue is in fact a practical one: when the three-sigma-plus black swan glides in, will there be enough liquidity provided by the less nimble to get out with a market order when your stop is triggered? BTW, your interest seems more than academic if you'll pardon the joke. Are you trading based on statistical analyses? Not asking for anything proprietary. Also, I think of the market as being conditionally distributed in three variables after you take on the position: a move against you, a move with you, and a consolidating neutral move. Those are the source statistics for a take profit rule, a stop loss rule, and a timeout rule. That make any sense to you? Best regards. - Mike
eSignal futures is $129/mo, DynaOrder $250 one-time fee, extensive historical intraday data ~$500... TS has it all for $100/mo. Commissions are equal to IB's now. Looks like it's the best deal.
Do you get a 100 dollar discount if you open a brokerage account? On the web site it says 199.5 per month. Are they adding exchange fees then?
I've seen several posts about "Wealthlab". What is Wealthlab and where can I find out more about it (BTW, www.wealthlab.com is a website under construction, and hence, has no info on it). Thanks in advance.
Volker, Yes. If you have a brokerage account with TS, the base platform fee is $99.95. If you exceed certain volumes (e.g., 250,000 share/month), then the base fee is waived for the next month. TS commissions are 1 cent for the first 500 shares (same as IB) and 0.6 cents above 500 shares (versus 0.5 for IB). TS charges you the SEC fee, whereas IB does not.