This is a good question!

1) You need to formalize and prepare your data. This involves

1.a) You need to understand how many time series you have

1.b) You need to understand how many variables relevant to sales data are available

1.c) You need to prepare a table with your source data. For example, it can contain the following columns:

product     date   sales    price   temperature

2) You need to formulate your research questions. What is the reason why you want to explore your sales data? Do you want to prepare a forecast or you want to analyse some dependencies?

3) You need to find appropriate models

4) You need to evaluate alternative models

5) You need to describe your results

Thank you very much , very clear explanation. 

I was wondering to ask how could we know how many time series we have ? I still did not catch it up ?

Thanks again 

I think in your case it depends on how many products you have.

How many products does your sales data relate to?

It is showed one type of product has been sold in different stores

Do you need to prepare forecasts for each of the stores?

Then the number of time series is equal to the number of stores, I guess.

No I need to forecast the future amount of sales - sum of stores sales for the incoming six weeks

so one time series for every week right ?

I think then you need to extrapolate one series.

How does your data look like? Is it weekly data? Do you have any additional variables? If you attach a graph, it would be easier to see what is needed to be done with the data.

its a daily data for three years long. Yes the data has variables like number of customers, holiday state and promotion .

I couldnt add the graph I will try one more

Thanks again and again

Ok - so you have daily observations, now some questions are

What type of forecast is needed? Do you need daily forecasts for the next six weeks? Do you need point forecasts only or do you need interval estimates?

Given that your sales are likely to be influenced by 1) promotions, 2) seasonal cycles, 3) moving state holidays, the task becomes non-trivial, 4) days of the week, and 5) unstable variance of modelling errors (volatility).

If you want to perform forecasting efficiently , you will need to build a model that takes into account all the above features of your data.

Firstly, you need to formalize your data - what is the info for promotions, what were the types of promotions, when they happened, what were the state holidays, etc.

When it comes to modelling, I think the best approach would be to build a SARIMA model with external regressors representing promotions and state holidays + seasonal cycles using trigonometric approximations. This is a well-known approach when you need to model early seasonal changes + daily changes + external factors. But you will probably need to apply Box-Cox transformations because you will need to stabilise error variance.

More on efficient modelling for forecasting with complex seasonality:

http://robjhyndman.com/papers/ComplexSeasonality.pdf

However, the SARIMA-based modelling and/or trigonometric approximation can be time-consuming.

If you have limited time resources, you can apply time series decomposition. You will need to extract trend and calculate seasonal indices first, then perform some analysis of the remaining part of the series, and then put everything back. I think more detailed explanations can be found in text books.

If you want something even more simple, try the Holt-Winters method, which is described in many text books and is easy to understand:

http://www.itl.nist.gov/div898/handbook/pmc/section4/pmc435.htm

When using the Holt-Winters method in your case, you will first need to scale your data to take into account the promotion factor and state holidays effects. Then proceed as described in text books and seasonal indices to model weekly cycles (if you do not have any weekly cycles, just use Holt's method or simple exponential smoothing)

After you find a good model, you will need to evaluate it using some error measures (such as MAE or MSE since you have only one series). You will not necessarily need MAPE or MdAPE since they are used for measuring accuracy across many series.

That's what you would need to do, in a few words...

P.S. Some useful info on how to forecast daily dales can be found here:

http://robjhyndman.com/hyndsight/dailydata/

http://stackoverflow.com/questions/22113819/r-holt-winters-with-daily-data-forecast-package

Thank you very much . I need your help , could you recommend a good book about decomposition time series forecasting ? I need to know more about . 

Again, thanks 

Start with this one: https://www.otexts.org/fpp/6

Read about classical decomposition.