I am looking for some thoughts on how best to approach a dataset and pick the right model for predicting the second hand price of widgets.
Some background on the properties of these widgets:
- The widgets come in different models, sizes, colors and a few other properties that are categorizable, along with the level of wear & tear that is unique for each widget (but can be measured accurately).
- The value of widgets depreciates over time (it depreciation profile is similar to 1/x) up until some fixed value which is equal to their scrap value. The scrap value is not constant over time but > 0.
- The widgets’ prices are observed once in a time period and recorded along with other properties for each widget. The price is not updated again.
- There is constant stream of new widgets hitting the market and the data comes pretty uniformly so the number of new widgets coming to the second hand market is pretty stable at say some k widgets per day s.t. the stock in time t+10 would be stock at time t plus 10*k.
Questions on how to handle the data:
- Because the price is observed only once, what should be the approach to setting upper limit as to how old the price data can be to be taken into consideration in the analysis? What kind of preliminary analysis should be used to determine such cutoffs?
- Also, should there be any weight put on the recency of the price data point in modeling the price?
How to choose the best model?
- What things should be considered about the underlying data before choosing the right model (treating outliers etc)?
- Is multiple variable regression the right way to approach this data or is there already an establish method of approaching such cases of price prediction?
