Bar based Autotrading

[Logic]

From what I'm gathering, this is what you're hypothesizing:

* A backtest using data that's 1 bar old outperforms a backtest using data that's 2 bars old

* Therefore, if going live where we can react to the current bar, it should outperform the 1-bar-back backtest in the same way the 1-bar-back backtest outperforms the 2-bar-back backtest.

Unfortunately, this isn't true. There is a step before going live, which is forward-testing - simulation in real-time. If you do that before going live, you'll be able to see the issues I'm going to present:

[Using 1-minute bars] When a bar forms, you have all sorts of actors moving on it; a bar that eventually turns into a "buy" bar might not do so until second #53 (for instance). That's not a problem, except it's possible that there was a very short-term sell-off from seconds #1 through #40. Look for Doji candles on the chart - they're the type of bar that cause these events the most.

If you were using a bar-based system to react against this real-time bar, your system would generate an erroneous sell signal for the first 40 seconds (depending on how sensitive your system is), and then nothing for the next 13 seconds, and then a buy signal for the last 7 seconds.

Since a bar system is only designed to react once per bar, your realtime behavior here would be to generate the sell signal. This is "unexpected" behavior, since the true action your strategy is supposed to perform is a buy signal - since bar systems analyze once-per-bar-at-the-end-of-the-bar.

Another scenario is that your bar system erroneously behaves like a tick system, generating a sell signal first, and then a buy signal later in the same bar, causing you to wash out of your entry for a loss.

I hope I've interpreted your question properly.

 

[walterjennings]

Quote from Logic:

I hope I've interpreted your question properly.

Yup. I agree there are situations where recalculating based on the newest bar may be bad. The question is, on average, will it have a positive effect or not, my price calculations are fairly slow and normally only shift around 1-2 cents per 1 min bar. a spike up in the bar may cause the price levels to shift up before the level is it, which would result in a more adventurous fill (or miss). I am unsure if the system 'dragging its feet' or being more reactive will get better performance, the only way for me to know for sure is to try trading them in separate accounts at the same time. Though I was hoping someone on here would have tested this theory live before.

 

[Algo_Design_Kid]

Quote from walterjennings:

Yup. I agree there are situations where recalculating based on the newest bar may be bad. The question is, on average, will it have a positive effect or not, my price calculations are fairly slow and normally only shift around 1-2 cents per 1 min bar. a spike up in the bar may cause the price levels to shift up before the level is it, which would result in a more adventurous fill (or miss). I am unsure if the system 'dragging its feet' or being more reactive will get better performance, the only way for me to know for sure is to try trading them in separate accounts at the same time. Though I was hoping someone on here would have tested this theory live before.

Logic is pretty correct here on the post before yours.

I have tried what you are trying to do. I have always used something like this.


Ex.

= AND (Buy Limit @ Previous Bar Close - .5 && Limit > Previous Bar Low)

The problem you will run into when forward testing live is another problem as well. You will never get the same fills on a live forward testing SIM as you will when you are actually running live with money on the line (Non-SIM forward testing).

If you continue on this path of wanting to use these methods I seriously recommend using longer term bars for testing (15 min ++).

Honestly, I kind of think you are wasting you time with anything less, but if nothing else at least you will learn from what you are doing. That is probably the most important part.

Depending on the software provider you are using as well you could really have some messed up non-real world fills going on. That I found over the past few years is the most detrimental part of this whole process. It becomes quite depressing, trust me, I have gone through the emotional ringer so to speak with this entire process.

GL

 

[borland]

With backtesting, what you want to see is results that are 'representive' of forward testing or actual trading.

There are many factors that can change actual results. But a representative sample is all you need, as future results are never predictable. All you can hope for is trade expectancy.

On my system using StrategyDesk, I have it set to 'buy the open' on the 5 minute bar. So my backtesting is done also on 'buy the open' by always looking back at the most recent closing bar before calculating various indicator values in the trade trigger logic.

Since my system is multi-timeframe, the backtester must also be set to 'calculate intra-bar on 5 minute intervals' for longer time periods than my systems 5 minute primary bar. Otherwise, the backtester would be calculating indicator values by looking forward to the close of those longer interval indicator values, there by providing un-representive results.

Hope your system's backtester has these same features.

 

[borland]

With backtesting, what you want to see is results that are 'representive' of forward testing or actual trading.

There are many factors that can change actual results. But a representative sample is all you need, as future results are never predictable. All you can hope for is trade expectancy.

On my system using StrategyDesk, I have live trading or forward testing set to 'buy the open' on the 5 minute bar. So my backtesting is done also on 'buy the open' by always looking back at the most recent closing bar before calculating various indicator values in the trade trigger logic.

Since my system is multi-timeframe, the backtester must also be set to 'calculate intra-bar on 5 minute intervals' for longer time periods than my system's 5 minute primary bar. Otherwise, the backtester would be calculating indicator values by looking forward to the close of those longer interval indicator values, there by providing un-representive results.

Hope your system's backtester has these same features.

 

[borland]

To clearify I actually meant to say "I have coded my trade logic to 'buy the open' by looking back at the prior bar". So, there are no trade alert settings for 'buy the open' in StrategyDesk. StrategyDesk only has this option in the backtester.

 

[Eight]

Quote from walterjennings:

Unfortunately I simulate my systems in indicators. So looking inside bar strategy simulations wouldn't work for me. I guess the question is, if you have a system that is profitable running levels at bar X-1, slightly less profitable running at bar X-2. If you were to move the level calculation to bar X (the live bar), would you expect to see a further performance increase (during live trading, not backtesting).

So as far as backtesting is concerned. I am not interested in testing at a tick level. Merely interested in the change in forward live testing performance change when running a back tested bar system on the live bar.

Tradestation will treat the current bar the same as in realtime. It will calculate the indicators to the tick, every tick...

 

[trade4ever2day]

As long as you make reference only to the O of a bar in your calculations and no reference is made to C, H or L, then you can use current bar.

The following is forward-looking

If current bar H > previous bar H

because the value of current bar H may change after you get a signal and your position will be invalid.

I hope this has relevance to your question.

 

[jack hershey]

Quote from walterjennings:

Sorry, I am horrible at explaining myself sometimes. I understand it would give me faster reacting levels + more accurate back tests if I converted all my systems over to Tick level. I am not interested in investing the time and money at this point and prefer to just discuss the idea of improving the current bar based system's real life performance, ignoring the limitations of back testing.

I am looking for intelligent insight on the following hypothesis from people who have experience trying this:

Hypothesis H: Systems trading using newer (more relevant data/information) will generally out perform systems which trade using more outdated data/information.

If system A ran on the newest bar X, using price levels from bar X-1, produces positive profit curve P(A,X,X-1).


We are unable to use the back test profit curve of P(A,X,X) to gauge its potential. This is because trading on newest bar X using price levels calculated from newest bar X gives unobtainable performance, since that allows the system to 'see into the future', since trades happen between the end of X-1 and the end of X, but the price levels get generated at the end of X.

We notice that we obtain a less desirable profit curve from running P(A,X,X-2), which would lend to the idea hypothesis H.

The question becomes, even though we can no longer back test P(A,X,X), would we expect to see a similar improvement in profit curve (running live) that we saw from moving from P(A,X,X-2) to P(A,X,X-1)? Where P(A,X,X) continuously updates trade levels intrabar based on the newest information.

Yes, you CAN get marginally better results.

Marginal improvement can also be significant.

Please think harder about your opportunity.

 

[abattia]

Quote from alexandermerwe:

... If current bar H > previous bar H ...

Actually, as soon as this condition becomes true in the current bar, it will remain true for the rest of the current bar, regardless of what else happens.

No?