CODEWITHSUNDEEP
machine-learningtime-seriesregression
Ali Tahir
Ali Tahir

Reputation: 3

How to decide a predictive model for sales forecasting

I would like to know which model should I choose to forecast monthly sales. should I go for regression approaches or time-series methods for small 1.5-year data?

Upvotes: 0

Views: 251

Answers (2)

Timbus Calin
Timbus Calin

Reputation: 15023

One of the first steps I would make is to clearly determine how many features you have.

In case of Univariate forecasting (observations in time of a single variable), you would most likely resort to even statistical approaches, such as ARIMA/SARIMA(I assume the concept of seasonality is known; if not, please read on properties of time series here : https://www.dummies.com/programming/big-data/data-science/key-properties-of-a-time-series-in-data-analysis/.

If you have multiple features(observations in time of multiple variables), you could first try with a VAR(vector autoregression).

Try these models at first, and only then proceed to more complicated ones such as LSTM/CNNs

Supporting @Nicolae Petridean's affirmation, the principle of Occam's Razor should always be applied: start with simple models and only after having tried several simpler ones should you progress to deep learning techniques.

Also, bear in mind that in the case of the latter, you will need much more data as compared to simpler statistical/mathematical models or even classical machine learning ones.

Upvotes: 1

Nicolae Petridean
Nicolae Petridean

Reputation: 604

Depending on the data that you have either one or the other might work. Or other techniques. Try 2 simple models using each of the 2 techniques, and validate them against a common validation dataset. This way you will have your answer. Nobody can answer to your question unless has quite some good insights into the data that you have for training. Out of my belly I would probably start with a regression but in the end I assume you will end up using something else. It is always a good option to start with simple models first to better understand the problem and then progressively fine tune or do other tricks and more complicated models, depending on what the models you already have learn or not.

Have a look at this Kaggle competition : https://www.kaggle.com/c/competitive-data-science-predict-future-sales

Check several notebooks from there and maybe you will understand more on what works or does not work in this kind of prediction.

Link to notebooks : https://www.kaggle.com/c/competitive-data-science-predict-future-sales/notebooks

Upvotes: 0

Related Questions