055 DDG Large Destroyer Thread

[AndrewS]

The figures used by the CSBA study are basically nonsense because their methodology is extremely flawed.

The CSBA assumption is that the Chinese cost structure is equivalent to the American one.

But if you're comparing relative costs of warships and aircraft, there's a big difference.

Chinese aircraft costs are probably approaching US levels, but in terms of surface warships, I see many examples where you can buy 2 or 3 Chinese warships for the equivalent of a single US warship. Part of the reason is that China doesn't have the same scale as the US in aerospace. In shipbuilding, the US doesn't have a civilian industry, whereas China is the world's largest shipbuilder.

The USAF also officially stated that an equivalent hypersonic missile costs 20x less in China.

 

[Staedler]

I don't want to spend hours explaining all the ways this report is problematic so I'll just start off what you would get in a first impression.

The first thing that jumps out at you when you skim the report is this graph they use. These are linear formulas on non-complex independent variables. Using linear regression like this to explain complex behavior is an immediate red flag.


So you take a look at their blurb explaining their choices.

In aggregate, we treated the outputs of these models as competing expert opinions, though
we favored linear regression since that method produces outputs that are more explain-
able than advanced machine learning methods.72

72 For the same reasons, linear regression was the favored cost estimation method for most platform types, where
possible. Although linear regression was the preferred cost estimating method, the final cost estimates used in the
SCT reflect the authors’ holistic assessment of the outputs of the various cost estimating methods.

So they picked linear regression despite the issues it has because it is more "explainable". And then they say adjusted it to reflect "holistic assessment" which is another way of saying they pulled it out of their ass.

The report is filled with stuff like this where they are constantly trying to pull a fast one on you. Run this by any data scientist and they would tell you it's a load of bull.

By the way, the formula for the trend lines on the above two graphs are:
Aircraft Cost = $9E-4 * MTOW + 66
Aircraft Cost = $2E-3 * Thrust + 26
Try plugging some known values in there like for B-2 or F-15. The prediction is so far off it's basically nonsense.

These are the sorts of formulas they are using to predict US costs. So if their first step has this much error in it, how much error do you think is in the final result when they try to use it to make relative cost predictions of the Chinese military?


I wasn't that negative about the study when it came out because it seemed like a first step and could potentially lead to something actually useful down the road. That's what the authors of the study seem to think too. But a first step is still a first step on a hundred mile journey. Yet I keep seeing people trying to use this study as an actual data point in discussion. With the level of inaccuracy in this study's results, the actual relative cost could even be 30 J20s to a single Type 052D. I can't stress enough that the error bars here are enormous. There is basically no useful information that can be gleaned from the report.

 

[Zichan]

I don't want to spend hours explaining all the ways this report is problematic so I'll just start off what you would get in a first impression.

The first thing that jumps out at you when you skim the report is this graph they use. These are linear formulas on non-complex independent variables. Using linear regression like this to explain complex behavior is an immediate red flag.

View attachment 96269
So you take a look at their blurb explaining their choices.


So they picked linear regression despite the issues it has because it is more "explainable". And then they say adjusted it to reflect "holistic assessment" which is another way of saying they pulled it out of their ass.

The report is filled with stuff like this where they are constantly trying to pull a fast one on you. Run this by any data scientist and they would tell you it's a load of bull.

By the way, the formula for the trend lines on the above two graphs are:
Aircraft Cost = $9E-4 * MTOW + 66
Aircraft Cost = $2E-3 * Thrust + 26
Try plugging some known values in there like for B-2 or F-15. The prediction is so far off it's basically nonsense.

These are the sorts of formulas they are using to predict US costs. So if their first step has this much error in it, how much error do you think is in the final result when they try to use it to make relative cost predictions of the Chinese military?


I wasn't that negative about the study when it came out because it seemed like a first step and could potentially lead to something actually useful down the road. That's what the authors of the study seem to think too. But a first step is still a first step on a hundred mile journey. Yet I keep seeing people trying to use this study as an actual data point in discussion. With the level of inaccuracy in this study's results, the actual relative cost could even be 30 J20s to a single Type 052D. I can't stress enough that the error bars here are enormous. There is basically no useful information that can be gleaned from the report.

To me the bigger issue appears to be the small size of the data sample, rather than linear regression per se.

I’m not actually convinced by your argument that linear regression is at fault here. You could argue that their single variable linear function model is too simplistic, but linear regression can be applied to multiple variable functions, polynomials, etc. Although, to be honest I expected that they would fit a multivariable function with all relevant platform parameters against procurement cost.

Then there’s also the fact that shipbuilding and aerospace are not comparatively competitive industries in the US and China. Shipbuilding in China is world class, whereas the US lags both in technology and even more so in scale.

When it comes to aerospace the roles are reversed, with the US enjoying both technological and scale leadership.

 

[Staedler]

To me the bigger issue appears to be the small size of the data sample, rather than linear regression per se.

I’m not actually convinced by your argument that linear regression is at fault here. You could argue that their single variable linear function model is too simplistic, but linear regression can be applied to multiple variable functions, polynomials, etc. Although, to be honest I expected that they would fit a multivariable function with all relevant platform parameters against procurement cost.

Then there’s also the fact that shipbuilding and aerospace are not comparatively competitive industries in the US and China. Shipbuilding in China is world class, whereas the US lags both in technology and even more so in scale.

When it comes to aerospace the roles are reversed, with the US enjoying both technological and scale leadership.

Yes, I've mentioned the small sample size as well as industry state comparisons before in another criticism. I just wanted to zoom in on one detail in my post which I haven't gone over before instead of spending time going every single issue I have with the study. In general I find the entire thing too flawed to be useful in any discussion.


When it was first brought up over at "PLA News, Pics and Discussion"

I was looking through that article trying to figure out how they estimate these costs and I'm not impressed.
For example they claim an R-squared adjusted of 0.92 for predicting fixed-wing aircraft but the shown graphs appear to have large residuals and the sample sizes are tiny. They appear be using sample sizes of around 17-20 which doesn't inspire confidence on it's predictive power.

They acknowledge the significant differences between platforms of US and Chinese weapons, but handwave it away in the next sentence by saying they picked the closest physical US analog for their cost analogy. And according to them these analogs were picked based on physical characteristics, no concern given for different production years or industry states.

Further reading yields more of the same. No adjustment made for the vastly different industrial states of the US and China. For example, I would expect the Chinese navy to be significantly cheaper in relation to the other branches than the balance that exists in the US since the US no longer has any civilian shipbuilding. Nor the relative cost differences that must be being imparted by advances in manufacturing techniques such as those in solid-fueled rockets. It's all handwaved away.

Using relative costs is an interesting idea, but the deficiencies in the model and results yields very little of the desired insight into PLA modernization trajectory. Actually tells me more about the US's trajectory than anything else.

Not going to even wade into whether or not their sources are worthwhile.

 

[AndrewS]

With the level of inaccuracy in this study's results, the actual relative cost could even be 30 J20s to a single Type 052D. I can't stress enough that the error bars here are enormous. There is basically no useful information that can be gleaned from the report.

The relative cost is actually the other way around.

My guestimate is around 3-4 J-20 for a single Type-052D.

 

[latenlazy]

To me the bigger issue appears to be the small size of the data sample, rather than linear regression per se.

I’m not actually convinced by your argument that linear regression is at fault here. You could argue that their single variable linear function model is too simplistic, but linear regression can be applied to multiple variable functions, polynomials, etc. Although, to be honest I expected that they would fit a multivariable function with all relevant platform parameters against procurement cost.

Then there’s also the fact that shipbuilding and aerospace are not comparatively competitive industries in the US and China. Shipbuilding in China is world class, whereas the US lags both in technology and even more so in scale.

When it comes to aerospace the roles are reversed, with the US enjoying both technological and scale leadership.

1. No amount of extra data will correct for how scattered the data already is. Variance is so high here with present data that even if you threw in say 10 more datapoints they’d have to be hugging the line of regression to tighten the r-square. And if you threw in 100 datapoints that hug the line you’d end up with a high count of outliers that would then need its own explaining.

2. The issue is not whether they used a multivariable or single variable function. The issue is that their choices of independent variables is nonsense. They’ve done the equivalent of regressing iPhone costs to its camera megapixel count.

The US does have a much bigger aerospace industry than China, but bigger industry alone doesn’t always translate to cheaper costs.

 

[kentchang]

The relative cost is actually the other way around.

My guestimate is around 3-4 J-20 for a single Type-052D.

Your US$500 million estimate sounds more reasonable. Between the two, a large ship is much better for force projection purpose (until a situation gets hot) as it can loiter for extended periods of time and has much greater range.

 

[YISOW]

Your US$500 million estimate sounds more reasonable. Between the two, a large ship is much better for force projection purpose (until a situation gets hot) as it can loiter for extended periods of time and has much greater range.

Thats ture
A single DD is better than 4 J-20

 

[davidau]

Thats ture
A single DD is better than 4 J-20

They have different role!

 

[Strangelove]

106 Yanan 延安