- Thread starter
- #1
If you are a regular follower of the Japan charts, you may have heard about COMG! pre-order (and order) points. To summarize, COMG! is a chain of 14 video game stores in the Niigata prefecture of Japan. They daily update the Top 20 pre-orders list and once a week, they publish the Top 20 best-selling games in their stores. As COMG publishes the sales on monday, impatient charts guys (like me) use that to estimate the Famitsu sales (published on thursday/friday) for years.
Common habit is to use a 1000 multiplier to assess the Famitsu first week sales (i.e. 50 pts COMG! sales => 50k Famitsu first week sales). But we remark some particularities and some variations depending on the type of game.
My objective is to build a predictive statistical model to estimate the ratio between COMG! points and Famitsu sales. Here is the tool :
How to use it ?
Just fill the info of the game you want to predict. Note that all info may not have an impact on the final prediction, that's normal (for example the day of the month does not do anything).
Methodology
The database gathers every NSW, PS4 and PS5 games that sold more than 50k first week according to famitsu and the corresponding COMG! info (orders and pre-orders) from 2017 to 2022. Since late 2022, I also keep track of every games ranked in both COMG! and Famitsu weekly sales. I can share the database with you on request (DM me).
Based on this, I applied a Random Forest with hardware, publisher, release date and of course COMG sales as variables.
How far is it from the truth ?
Of course, this prediction is an estimation, based on what we observed in the past and my modeling. Do not expect it to be 100% reliable (but I hope it won't be too far either^^). I will try to continue to improve it.
Note that by the end of 2023, the test error on the dataset is <35% , with significant variation from 0% to 80%.
Some improvement to come ?
To conclude
I hope you will find this tool useful and interesting. I just want it to be another input to the weekly debate/discussion about Japan video game sales that we all like. I you have any suggestion, if you have some data you want or you want to share with me, if you want more info about the statistical model just ask me
. This work was possible thanks to a lot of people (that are in the aknowledgement of the app), from VgChartz, Twitter or Installbase.
UPDATE from 01-25-2024
I have created a model based on monthly Amazon data. It is a statistical/machine learning model based on the Amazon sales at the end of last month, the number of days that are not included (comparison between the end of last month and the release date) and the caracteristics of the game (console, editor). It is an alternative model built te takle COMG model issues. For now, it is only on test (as we only have data from 08/2024, thanks to nichebarrier.com) and is not added to the App yet.
The next direction are :
Common habit is to use a 1000 multiplier to assess the Famitsu first week sales (i.e. 50 pts COMG! sales => 50k Famitsu first week sales). But we remark some particularities and some variations depending on the type of game.
My objective is to build a predictive statistical model to estimate the ratio between COMG! points and Famitsu sales. Here is the tool :
How to use it ?
Just fill the info of the game you want to predict. Note that all info may not have an impact on the final prediction, that's normal (for example the day of the month does not do anything).
Methodology
The database gathers every NSW, PS4 and PS5 games that sold more than 50k first week according to famitsu and the corresponding COMG! info (orders and pre-orders) from 2017 to 2022. Since late 2022, I also keep track of every games ranked in both COMG! and Famitsu weekly sales. I can share the database with you on request (DM me).
Based on this, I applied a Random Forest with hardware, publisher, release date and of course COMG sales as variables.
How far is it from the truth ?
Of course, this prediction is an estimation, based on what we observed in the past and my modeling. Do not expect it to be 100% reliable (but I hope it won't be too far either^^). I will try to continue to improve it.
Note that by the end of 2023, the test error on the dataset is <35% , with significant variation from 0% to 80%.
Some improvement to come ?
- I continue to collect info every week to build a more reliable method
- My next goal is to about collector edition. For now, collector edition are considered as regular sales. But we can feel that the ratio of these type of sales is probably much lower that a regular version. I should take this into account in the future.
- Improve my model : try other statistical method. Maybe use some neural network to see what happen.
- Take into account pre-orders. I would love to use pre-order dynamic to predict the finale sales of a game. But that requires a lot of work (collecting data and modeling), so not for now.
To conclude
I hope you will find this tool useful and interesting. I just want it to be another input to the weekly debate/discussion about Japan video game sales that we all like. I you have any suggestion, if you have some data you want or you want to share with me, if you want more info about the statistical model just ask me
. This work was possible thanks to a lot of people (that are in the aknowledgement of the app), from VgChartz, Twitter or Installbase.UPDATE from 01-25-2024
I have created a model based on monthly Amazon data. It is a statistical/machine learning model based on the Amazon sales at the end of last month, the number of days that are not included (comparison between the end of last month and the release date) and the caracteristics of the game (console, editor). It is an alternative model built te takle COMG model issues. For now, it is only on test (as we only have data from 08/2024, thanks to nichebarrier.com) and is not added to the App yet.
The next direction are :
- build a meta-model to decide when we should use which model
- build a new model that combines both source of information.
Last edited:
