- Joined
- 13 June 2007
- Posts
- 838
- Reactions
- 136
Hi Gorilla --
The problem of finding high quality historical data is not unique to the Australian markets.
Even for US stocks, historical data that lists the actual prices and volumes as traded, unadjusted for splits, distributions, and dividends, is very expensive.
To remove survivor bias, the data should include delisted issues as well.
Once the data is located, the problem is further complicated Because the data is unadjusted. As most of you who have tried to clean your data know, it is often impossible to tell when a 10 or 20 or 50% price change is a price change or a split. The trading system software must be able to cope with the discontinuities, which few (none?) can. For comparison, think about the way trading system packages that are designed to work with futures contracts handle the roll-over from one front month to the next. They have an algorithm based on either a calendar or on volume that tells them when to switch, and they take the differential between the two contracts into account as they report the trading results. Neither of those methods can be applied to stock distributions.
In an ideal world: we would have perfectly clean data; there would be a field for Actual price along with Open, High, Low, Close; there would be a field for a code that identifies what happened at every discontinuity; and our development software would be able to handle all of this.
But that still does not enable us to handle situations where the fundamentals of the company change without the ticker changing and without a distribution. For example, an automobile manufacturer sells its financing division. Among US automakers, the financing division accounts for nearly all of the profit of the company. The profitability and consequent price action change, but there is no way to determine what happened without reading the press releases, and there is no way to adjust previous prices to remove the contribution of the financing division.
I do not think we can completely, or even adequately, test for the effect of historical actual prices, splits, dividends, takeovers, spinoffs, or survivor bias.
And I think it is Very dangerous to draw conclusions from long-only systems in the 1982 to 2007 time period.
Thanks for listening,
Howard
www.quantitativetradingsystems.com
Really good point Howard. It appears to mean that stats generated from testing a system on data flawed in the ways you describe must be dubious. But no-one seems to have ASX data that isn't adjusted for splits or with the dividends adjusted into the share price. I've seen the Blackstar study and agree it would be far nicer to have this kind of data to test with (not to mention a CPI adjusted liquidity filter...this is obviously far easier to do). Without this kind of data, what do you do? This could be yet another reason why the big boys don't play in this part of the market..the Aust market isn't big enough to warrant fixing 22 years of data for testing systems like what the Blackstar guys do in the US.
The problem of finding high quality historical data is not unique to the Australian markets.
Even for US stocks, historical data that lists the actual prices and volumes as traded, unadjusted for splits, distributions, and dividends, is very expensive.
To remove survivor bias, the data should include delisted issues as well.
Once the data is located, the problem is further complicated Because the data is unadjusted. As most of you who have tried to clean your data know, it is often impossible to tell when a 10 or 20 or 50% price change is a price change or a split. The trading system software must be able to cope with the discontinuities, which few (none?) can. For comparison, think about the way trading system packages that are designed to work with futures contracts handle the roll-over from one front month to the next. They have an algorithm based on either a calendar or on volume that tells them when to switch, and they take the differential between the two contracts into account as they report the trading results. Neither of those methods can be applied to stock distributions.
In an ideal world: we would have perfectly clean data; there would be a field for Actual price along with Open, High, Low, Close; there would be a field for a code that identifies what happened at every discontinuity; and our development software would be able to handle all of this.
But that still does not enable us to handle situations where the fundamentals of the company change without the ticker changing and without a distribution. For example, an automobile manufacturer sells its financing division. Among US automakers, the financing division accounts for nearly all of the profit of the company. The profitability and consequent price action change, but there is no way to determine what happened without reading the press releases, and there is no way to adjust previous prices to remove the contribution of the financing division.
I do not think we can completely, or even adequately, test for the effect of historical actual prices, splits, dividends, takeovers, spinoffs, or survivor bias.
And I think it is Very dangerous to draw conclusions from long-only systems in the 1982 to 2007 time period.
Thanks for listening,
Howard
www.quantitativetradingsystems.com