Financial AI Model Study
Compared regression, dimensionality-reduction, tree, support-vector, and neural-network approaches to understand when model complexity improves generalization.
Quick read
The through-line of the work.
Open
The starting question
How do preprocessing, regularization, model complexity, and hyperparameters affect validation and test performance?
Build
What I checked
I started with Four applied assignments using financial and public datasets and moved through Test model complexity -> Examine representation and regularization -> Tune support-vector models -> Compare neural-network depth.
Turn
What changed my view
The first two reported PCA components explained approximately 88.4% of variance in the national-risk example.
Close
Where I draw the line
Original datasets and data dictionaries should be recovered.
01 · Start
The question I began with
How do preprocessing, regularization, model complexity, and hyperparameters affect validation and test performance?
02 · Material
The data and evidence
03 · Work
How I worked through it
- 01
Test model complexity
Compared polynomial degrees and decision-tree settings using held-out performance.
- 02
Examine representation and regularization
Applied PCA, standardization, Ridge, and Lasso to understand structure and overfitting.
- 03
Tune support-vector models
Evaluated Linear SVR parameter combinations on Iowa housing data.
- 04
Compare neural-network depth
Compared linear regression, a one-hidden-layer ANN, and a three-hidden-layer DNN.
04 · Takeaway
What the numbers suggested
The first two reported PCA components explained approximately 88.4% of variance in the national-risk example.
Linear SVR with C=1 and epsilon=50 was strongest among the tested combinations.
Additional DNN depth produced little improvement over the smaller ANN.
Not claimed
- Does not claim applied business impact. These are assignment-level comparisons.
- Does not claim one model is always better. It depends on the data and the goal.
05 · Use
Where this helps
Use additional complexity only when validation evidence justifies it; a more advanced method is not automatically the stronger business choice.
06 · Boundary
What still needs checking
- •Original datasets and data dictionaries should be recovered.
- •The four assignments should be consolidated into reproducible notebooks.
- •Classroom metrics should not be presented as production performance.
07 · Figures
Tables and figures
Polynomial degree and validation RMSE
Reported evidenceThe bars reproduce validation RMSE values reported in assignment 1. They do not by themselves establish a preferred final degree; training and test behavior still need a reproducible notebook.
Source · Financial AI assignment 1 report
PCA explained variance
Reported evidencePC1 and PC2 explain 64.03% and 24.37%, about 88.4% combined in the reported national-risk example. This is classroom evidence for that dataset, not general financial-model performance.
Source · Financial AI assignment 1 report · national-risk example
Linear SVR tuning result
Reported evidenceC
1.0
epsilon
50
Train MSE
1,782.2306
Validation MSE
1,944.5111
Among the tested combinations, C=1 and epsilon=50 produced the lowest reported validation MSE. This is a bounded classroom experiment, not evidence of external or production performance.
Source · Financial AI assignment 3 report · Iowa housing example
ANN and DNN validation behavior
Reported evidenceANN
One hidden layer · comparison baseline
DNN
Three hidden layers · limited gain from added complexity
The report concludes that the deeper DNN did not deliver a meaningful improvement over the smaller ANN. Because minimum-loss figures are not fully consistent within the document, only the qualitative conclusion is shown.
Source · Financial AI assignment 4 report