Valentina Pancaldi
Data Analyst building ML models & forecasting — SQL, Python,
New to Contra
Valentina is ready for their next project!
ML Model to Predict Litigation Risk in Insurance Claims
Currently developing a machine learning model that estimates, for
each individual claim, the probability that it will end up in
litigation — based on features learned from historical claims data.
Following the same 5-stage process as our forecasting work: business
& data understanding, data assessment & cleaning, data preparation,
model development/evaluation/deployment, and implementation &
knowledge transfer.
This project is currently in progress.
0
6
Logistic Regression from Scratch for Pneumonia Detection in Chest X-Rays
Built a binary classification model from scratch (no ML libraries —
implemented gradient descent and the loss function manually) to detect
pneumonia from chest X-ray images, for a Numerical Methods course project.
Preprocessing: converted each X-ray into a 32x32 grayscale image,
flattened into a normalized vector, and balanced the training set to
have equal numbers of NORMAL and PNEUMONIA cases to avoid class bias.
Model: binary classifier using a tanh-based activation function,
trained via gradient descent with manually derived partial derivatives
(no autograd).
Process: tested 5 different learning rates (alpha = 0.0001 to 0.5) to
find the best trade-off between convergence speed and stability,
tracking MSE on both train and test sets across iterations.
0
6
Built a forecasting model to predict the number of insurance claims
expected over the next 6 months, broken down by business-relevant
categories, for a large-scale insurance claims dataset.
The project followed a full 5-stage process: business & data
understanding, data assessment & cleaning, data preparation, model
development/evaluation/deployment, and implementation & knowledge
transfer.
We tested and compared multiple forecasting approaches — Auto ARIMA,
Auto ETS, Auto Theta, Prophet, LightGBM, and XGBoost — to select the
best-performing model for each claim category.
Results:
Improved MAPE by 47%, from a 6.42% baseline to 3.43% in production
The production model outperformed the baseline on 96% of claim types
Identified and corrected 19+ data quality issues during the cleaning phase
Delivered 48 visualizations giving the client deeper insight into their own claims data
1
20
Data analyst working with SQL, Python, PySpark and Databricks —
currently building forecasting models on insurance claims data.
Open to new opportunities — freelance, contract, part-time or
full-time. If you need a hand cleaning a dataset, building a model,
or automating something with Python, let's talk.
0
13