From COMG! point to Famitsu sales: a predictive tool [NEW: Amazon Data is now used for the prediction]

[ChartsJapan]

Here are the forecasts for last week releases:

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED	Famitsu	Error
PS5	Like a Dragon Gaiden: The Man Who Erased His Name	Sega	2023.11.09	51​	60163​	63319​	-5%​
PS4	Like a Dragon Gaiden: The Man Who Erased His Name	Sega	2023.11.09	79​	62478​	60134​	4%​
PS5	Call of Duty: Modern Warfare III	Activision	2023.11.10	15​	21140​	22132​	-5%​
PS4	Call of Duty: Modern Warfare III	Activision	2023.11.10	7​	6504​	8962​	-38%​
PS5	Tales of Arise: Beyond the Dawn Edition	Bandai Namco	2023.11.09	3​	4472​	4279​	4%​
PS4	Tales of Arise: Beyond the Dawn Edition	Bandai Namco	2023.11.09	4​	3310​	1812​	45%​
NSW	Double Dragon Collection	Arc System Works	2023.11.09	3​	4636​	3094​	33%​

Cumul:
LaDG PS5/PS4: 122.004 (123.453, ERROR : -1%)
CoD PS5/PS4: 27.644(31.094, ERROR : -12%)
ToA PS5/PS4: 7.782(6.091, ERROR : 22%)

Seems ok overall.
I would have put LaDG PS5 > PS4 despite the big gap in COMG sales

EDIT : pretty good forecasts

 

[ChartsJapan]

Here are the forecasts for last week releases:

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED	Famitsu	Error
NSW	Super Mario RPG	Nintendo	2023.11.17	197​	222803​	301334​	-35%​
NSW	Momotaro Dentetsu World: Chikyuu wa Kibou de Mawatteru!	Konami	2023.11.16	156​	181240​	314699​	-74%​
NSW	Persona 5 Tactica	Atlus	2023.11.17	22​	28712​	26794​	7%​
PS5	Persona 5 Tactica	Atlus	2023.11.17	14​	19423​	15713​	19%​
PS4	Persona 5 Tactica	Atlus	2023.11.17	4​	3482​	7421​	-113%​
NSW	Hogwarts Legacy	WB Games	2023.11.14	17​	21363​	47717​	-123%​

Cumul:
Persona Tactica NSW/PS5/PS4: 51.617 (49.928, ERROR: 3%)

Seems low for HL. Momotaro is hard to predict.

EDIT : very bad forecasts. COMG was very low

 

[Arynio]

Here are the forecasts for last week releases:

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED
NSW	Super Mario RPG	Nintendo	2023.11.17	197​	222803​
NSW	Momotaro Dentetsu World: Chikyuu wa Kibou de Mawatteru!	Konami	2023.11.16	156​	181240​
NSW	Persona 5 Tactica	Atlus	2023.11.17	22​	28712​
PS5	Persona 5 Tactica	Atlus	2023.11.17	14​	19423​
PS4	Persona 5 Tactica	Atlus	2023.11.17	4​	3482​
NSW	Hogwarts Legacy	WB Games	2023.11.14	17​	21363​

Cumul:
Persona Tactica NSW/PS5/PS4: 51.617

Seems low for HL. Momotaro is hard to predict.

First time seeing this thread. Pretty interesting stuff. I believe SMRPG, Momotaro and HL will all be underestimates once again. P5T should be fine, I think. Can't wait to see the results.

 

[Aostia82]

I wonder if we could develop a "cumulative" predictive tool matching also Amazon actual sales for predictions

 

[ChartsJapan]

I wonder if we could develop a "cumulative" predictive tool matching also Amazon actual sales for predictions

That is exactly what I want to. I have different possibilités, but my favorite for the moment is :
-comg predictions with a model
-Amazon prediction with Another model
- a third model that should decide which model to use depending on the game caracteristics

The reason why I go for that, is that forecast quality depends in the game caracteristics (and may vary from a model to another). For example, PS5 games are already quite well predicted by my current Amazon model. Plus as I use nichebarrier Amazon data, i only have monthly data (which is already nice, tks to him for his work). But, that means I may have a lot of missing days, depending on the release date in the game. That is also an information that is useful to decide which model to use.

Another solution would be to combine both model predictions (e.g. With random Forest frame work).

If you have ideas, do not hesitate.

Anyway, we should wait to have Amazon more data. But I will probably propose a first hybrid model by the beginning of 2024

 

[Aostia82]

That is exactly what I want to. I have different possibilités, but my favorite for the moment is :
-comg predictions with a model
-Amazon prediction with Another model
- a third model that should decide which model to use depending on the game caracteristics

The reason why I go for that, is that forecast quality depends in the game caracteristics (and may vary from a model to another). For example, PS5 games are already quite well predicted by my current Amazon model. Plus as I use nichebarrier Amazon data, i only have monthly data (which is already nice, tks to him for his work). But, that means I may have a lot of missing days, depending on the release date in the game. That is also an information that is useful to decide which model to use.

Another solution would be to combine both model predictions (e.g. With random Forest frame work).

If you have ideas, do not hesitate.

Anyway, we should wait to have Amazon more data. But I will probably propose a first hybrid model by the beginning of 2024

Sounds great!!!
Thanks for the effort

 

[ChartsJapan]

Last main release of the year :

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED	Famitsu	Error
NSW	Dragon Quest Monsters: The Dark Prince	Square Enix	2023.12.01	213​	230138​	346583​	-51%​

This one will be particularly difficult as :

• The COMG contains a huge amount of Special Edition
• The game is sold out according to Square-Enix

EDIT : big fail ^^

 

[ChartsJapan]

Hello !
For 2024, I introduce a new model, based on monthly Amazon sales (from nichebarrier.com, thanks). We only have data from august 2023 for this one, but it already may provide some relevant info. Thus, I will post predictions of both models for a while, before trying to select the best one or to combine both.
Note that, as the Amazon sales are monthly sales, so if a game releases by the end of the month, it may be more difficult to predict.
We have 2 new releases this week, both scoring very low at COMG.

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
NSW	Another Code: Recollection	Nintendo	2024.01.19	2​	2821​	2050​	30058​	15741​	-82%​	48%​
PS5	The Last of Us Part II Remastered	SIE	2024.01.19	<2	<3085​	600​	7068​	4364​	<-29%​	38%​

Forecasts of both model are very different ! It will be interesting to see the results

EDIT : Real figures were in the middle of both algorithm. I would still give the win for amazon.

COMG: 0
AMAZON: 2

 

[J-Cat]

Hello !
For 2024, I introduce a new model, based on monthly Amazon sales (from nichebarrier.com, thanks). We only have data from august 2023 for this one, but it already may provide some relevant info. Thus, I will post predictions of both models for a while, before trying to select the best one or to combine both.
Note that, as the Amazon sales are monthly sales, so if a game releases by the end of the month, it may be more difficult to predict.
We have 2 new releases this week, both scoring very low at COMG.

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON
NSW	Another Code: Recollection	Nintendo	2024.01.19	2​	2821​	2050​	30058​
PS5	The Last of Us Part II Remastered	SIE	2024.01.19	<2	<3085	600​	7068​

Forecasts of both model are very different ! It will be interesting to see the results

OO awesome!
Also you can check Amazon right now to see what Another Code sold in the past month 2300. Niche Barrier has collated the sales for Another Another up until Dec 29th 2023, so there's some overlap but one can add those together to get 4350. I usually times by 5 since Amazon is around 20% of the the physical market in Japan and that ends with a predication of 21,750 for Another Code. Your multiplier is a lot more generous at around 14.66 (I'm not sure how you are calculating it to be honest so I just divided 30,058 by 2050) which gives 63,781 which seems way too high lol.

Doing the same for The Last of Part II Remastered (Niche Barrier data through to Dec 29th + current Amazon sales X 5) = 8250
Using your multiplier (11.78) = 19,437

I'd be very interest to know how you made the Amazon prediction . Mine is crude lol. It worked when I used it for Raincode though since COMG was way off for that title.

 

[Aostia82]

Awesome! I remember we briefly discussed about using Amazon over Comg so this is an exciting test!

 

[ChartsJapan]

OO awesome!
Also you can check Amazon right now to see what Another Code sold in the past month 2300. Niche Barrier has collated the sales for Another Another up until Dec 29th 2023, so there's some overlap but one can add those together to get 4350. I usually times by 5 since Amazon is around 20% of the the physical market in Japan and that ends with a predication of 21,750 for Another Code. Your multiplier is a lot more generous at around 14.66 (I'm not sure how you are calculating it to be honest so I just divided 30,058 by 2050) which gives 63,781 which seems way too high lol.

Doing the same for The Last of Part II Remastered (Niche Barrier data through to Dec 29th + current Amazon sales X 5) = 8250
Using your multiplier (11.78) = 19,437

I'd be very interest to know how you made the Amazon prediction . Mine is crude lol. It worked when I used it for Raincode though since COMG was way off for that title.

Hey,
Thanks for the info about de Amazon sales. It will stick to my monthly sales, as it is faster and easier for me.
Concerning your method for multiplier, it seems pretty relevant. On my side, I build 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 provides me an estimator for the multiplier. As we lack of data, this model should improve with time and we should be patients ^^

 

[Lelouch0612]

You rock @ChartsJapan !

Updating the thread's title

 

[J-Cat]

Hey,
Thanks for the info about de Amazon sales. It will stick to my monthly sales, as it is faster and easier for me.
Concerning your method for multiplier, it seems pretty relevant. On my side, I build 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 provides me an estimator for the multiplier. As we lack of data, this model should improve with time and we should be patients ^^

Thanks for the explanation! Yeah I've just been roughly trying to work out what games would debut at on the Famitsu chart and I wasn't really sure how best to do it so that's what I came up with lol. I had no idea you was working on a method! I look forward to seeing the results.

 

[ChartsJapan]

You rock @ChartsJapan !

Updating the thread's title

Thanks, I will update the OP as well.

 

[ChartsJapan]

Here are the forecasts for last week releases. Vey busy week !

PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
Sega	2024.01.26	60​	73067​	10800​	123678​	102940​	-29%​	17%​
Sega	2024.01.26	77​	59016​	4600​	55454​	77134​	-23%​	-39%​
Spike Chunsoft	2024.01.25	45​	43562​	5400​	69791​	85424​	-49%​	-22%​
Bandai Namco	2024.01.26	20​	24444​	2450​	28854​	20516​	19%​	29%​
Capcom	2024.01.25	5​	5312​	2200​	28433​	10593​	-50%​	63%​

Cumul:

• LaDIW PS5/PS4:
  • COMG: 132.083 (FAMITSU 180.074)
  • AMAZON: 179.132 (FAMITSU 180.074, -1%)

Amazon forecasts are usually above the COMG ones. Great difference for LaD PS5, Shiren and Apollo Justice. More close predictions for the others.
Interesting week for sure!

EDIT : quite a decent week, except for Ace Attornay.

COMG: 3
AMAZON: 5

 

[ChartsJapan]

Hey guys,
I am late to post the performances of the algorithms for last week releases:

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
PS5	Persona 3 Reload	Atlus	2024.02.02	78​	102355​	20150​	217771​	76368​	34%​	185%​
PS4	Persona 3 Reload	Atlus	2024.02.02	46​	44410​	8950​	103581​	40024​	11%​	159%​
PS5	Granblue Fantasy: Relink	Cygames	2024.02.01	38​	54409​	11300​	123279​	48754​	12%​	153%​
PS4	Granblue Fantasy: Relink	Cygames	2024.02.01	23​	22283​	3350​	39151​	22085​	1%​	77%​
NSW	Jujutsu Kaisen: Cursed Clash	Bandai Namco	2024.02.01	23​	24921​	3850​	55289​	25242​	-1%​	119%​
PS5	Jujutsu Kaisen: Cursed Clash	Bandai Namco	2024.02.01	5​	7546​	800​	9129​	4009​	88%​	128%​

Amazon model prediction are completely off this week, with a systematical overestimation. Low Amazon ratio this week but still. Again, be patient, we only have a few months of background for amazon data (vs many years for COMG)
COMG model did pretty good.

The combined model (COMG and Amazon in one model) I am building is improving week after week as well. But for now, it is not performant enough to be part of the comparison (test error of 60% for combined vs 40% for COMG and Amazon)

TOTAL COMPARISON
COMG: 9
AMAZON: 5

 

[ChartsJapan]

Here are the forecasts for last week releases:

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
NSW	Mario VS. Donkey Kong	Nintendo	2024.02.16	18​	24033​	2750​	24531​	61930​	-61%​	-60%​
PS5	Cyberpunk 2077: Ultimate Edition	Spike Chunsoft	2024.02.15	4​	5936​	900​	6417​	3844​	54%​	67%​
NSW	Front Mission 2: Remake	Rainy Frog	2024.02.15	4​	5279​	350​	4634​	3182​	66%​	46%​
NSW	The Legend of Heroes: Trails through Daybreak	Nihon Falcom	2024.02.15	3​	4179​	600​	5541​	7160​	-42%​	-23%​
PS5	Skull and Bones	Ubisoft	2024.02.16	2​	3198​	11500​	77117​	6635​	-52%​	1062%​

Once again, amazon predictions are relatively high. Also note that the case Skull and Bones is very particular: the amazon sales are very high (11500) compared to COMG (2) but most of it were probably a long time ago (and may have been canceled since). Depending on the actual sales, I would probably remove S&B from the database.

EDIT 1: I found some bug in my code, which explains such a big overestimation from Amazon model for a few weeks. I have changed this week predictions and decided to reset the competition score between both model.
EDIT 2: MvsDK forecast were far from reality (but the reality was quite surprising^^). I will probably remove MvsDK and S&B from data base

TOTAL COMPARISON (start from skratch)
COMG: 2
AMAZON: 3

 

[ChartsJapan]

Here are the forecasts for last week releases:

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
NSW	Shin-chan: Shiro of Coal Town	Neos	2024.02.22	13​	16083​	1650​	17133​	21593​	-26%​	-21%​
NSW	Retro Game Challenge 1 + 2 Replay	Bandai Namco	2024.02.22	14​	16498​	3600​	39797​	13686​	21%​	191%​

Both models agree on Shin-Chan
Big amazon pre sales for RGC that leads to higher predictions.

EDIT : amazon sales of RGC were indeed to high

TOTAL COMPARISON
COMG: 3
AMAZON: 4

 

[ChartsJapan]

Here are the forecasts for last week releases, with notably FF VIIR!

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
PS5	Final Fantasy VII Rebirth	Square Enix	2024.02.29	182​	206580​	69900​	423910​	262656​	-21%​	61%​
NSW	Bart Stella Abyss	Nippon Ichi Software	2024.02.29	7​	8974​	300​	3740​	2077​	332%​	80%​
NSW	Goblin Slayer Another Adventurer: Nightmare Feast	Bushiroad	2024.02.29	4​	5250​	550​	5256​	6822​	-23%​	-23%​

Big difference for FFVIIR between COMG and Amazon, based on the different sales data range. Reality will probably be between these two.
For the two other games, Amazon prediction seem to be more trustworthy

EDIT : COMG was right for FFVIIR !

TOTAL COMPARISON
COMG: 4
AMAZON: 5

 

[ChartsJapan]

Here are the forecasts for last week release,

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
NSW	Unicorn Overlord	Atlus	2024.03.08	30​	35214​	8800​	59963​	40991​	-14%​	46%​
PS5	Unicorn Overlord	Atlus	2024.03.08	29​	35138​	5600​	31027​	24398​	44%​	27%​
PS4	Unicorn Overlord	Atlus	2024.03.08	10​	10293​	1050​	7795​	8797​	17%​	-11%​
NSW	Fitness Boxing feat. Hatsune Miku: Isshoni Exercise	Imagineer	2024.03.07	3​	4797​	4250​	30614​	14128​	-66%​	117%​

UO NSW seems to be sold out at COMG, so Amazon pred may be more realistic
Fitness Boxing probably overestimated by Amazon and underestimated by COMG.

It is in these cases, that the combined model, I am currently developping, should be more relevant. it will trade off between both predictions

EDIT : decent results this week

TOTAL COMPARISON
COMG: 6
AMAZON: 7

 

[ChartsJapan]

Here are the forecasts for the many releases of last week,

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
PS5	Dragon's Dogma 2	Capcom	2024.03.22	54​	65943​	9000​	67428​	68592​	-4%​	-2%​
PS5	Rise of the Ronin	SIE	2024.03.22	44​	56104​	6050​	47242​	64656​	-13%​	-27%​
NSW	Princess Peach: Showtime!	Nintendo	2024.03.22	35​	46958​	3600​	39286​	77562​	-39%​	-49%​
NSW	Dragon Quest X: Door to the Future and the Sleeping Girl Online	Square Enix	2024.03.21	4​	5368​	1450​	13376​	17919​	-70%​	-25%​
PS4	Dragon Quest X: Door to the Future and the Sleeping Girl Online	Square Enix	2024.03.21	3​	3342​	550​	5594​	7333​	-54%​	-24%​

Amazon forecasts seem more accurate for me, except maybe for Princess Peach.

EDIT : correct forecast overall, except for PPS which was largely underestimated

TOTAL COMPARISON
COMG: 8
AMAZON: 10

 

[ChartsJapan]

Here are the forecasts for the Winning Post 10 2024 versions:

PLATFORM	TITLE	PUBLISHER	RELEASE_DATE	SALES_COMG	PRED_COMG	SALES_AMAZON	PRED_AMAZON	Famitsu	Error_COMG	Error_AMAZON
NSW	Winning Post 10 2024	Koei Tecmo	2024.03.28	12​	16824​	1150​	8950​	12845​	31%​	-30%​
PS5	Winning Post 10 2024	Koei Tecmo	2024.03.28	2​	4186​	750​	5205​	6733​	-38%​	-23%​
PS4	Winning Post 10 2024	Koei Tecmo	2024.03.28	3​	5092​	350​	3194​	4702​	8%​	-32%​

Expect soft drop compared to last versions from 2023 (NSW 13k , PS5 7k, PS4 6k)

TOTAL COMPARISON
COMG: 9
AMAZON: 12