[ProgSoc] Predict Sydney traffic?
marauder at marauder.net.au
Fri Dec 3 00:17:57 EST 2010
Le 02/12/2010 à 9:55 PM, John Elliot a écrit :
> In broad strokes how would you go about solving a problem like this?
It probably won't win you any prizes in this competition, but the first hammer to hit that nail with is ARIMA. The Wikipedia article on ARIMA is pretty rubbish but their article on Box-Jenkins [https://secure.wikimedia.org/wikipedia/en/wiki/Box-Jenkins] is a decent summary of what's involved in making forecasts of something given only previous measurements of that thing.
More generally we say that time series are made up of four kinds of thing stacked on top of each other: trends, cycles, seasonality, and irregular movements. Easiest to define by example, consider monthly financial data. The trend is the long-term progression of the thing, not cyclical, emerging over periods like 15 years. Seasonality is the monthly fluctuations that recur every year: big Xmas, January slump, etc. Cycles are repeated fluctuations that aren't seasonal or regular and emerge over periods like 2--5 years. Irregular movements are point events that have temporary or permanent effects on the cycle: pulses, step changes, etc. You detect them to get rid of them, then fit curves to the trend, seasonality, and cycles so you can extrapolate.
Your next move is to get extra data such as weather that might have a causal relationship with your time series. ARIMA can take that into account too.
More information about the Progsoc