Harideep Aepuri

This practicum project aims at creating a machine learning regression model that can predict the Airbnb rental price based on publicly available listing information. The project was performed in several steps, such as data collection and preprocessing, exploratory data analysis, feature engineering and feature selection, model development and training, and model assessment and performance adjustment. All the stages have led to the construction of a total predictive model that could examine intricate pricing trends.

This study compared five time series forecasting models for predicting Bitcoin (BTC/USDT) close price at one-minute intervals. One-minute open-high-low-close-volume (OHLCV) data spanning 60 days from January 30 to March 31, 2026 was collected from the Binance exchange API, producing 86,341 records. The dataset was divided into a 45-day training set and a 15-day true out-of-sample test set. Five models were implemented and evaluated under a consistent rolling one-step-ahead protocol: ARIMA(2,1,2), Facebook Prophet, LSTM Univariate, LSTM Multivariate, and a Hybrid ARIMA+LSTM model. Feature engineering produced 43 technical indicators from raw OHLCV data, and XGBoost feature importance analysis reduced the multivariate LSTM input to 20 features. Evaluation used four metrics: mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and directional accuracy. Additional analyses included walk-forward cross-validation with four expanding folds, multi-horizon forecasting at H = 1,5,10,30 minutes, an extended 3-hour context window LSTM experiment (W = 180), multi-day forecast stability tests, and a cross-market study that incorporated Gold, NASDAQ, and the US Dollar Index as potential predictors for daily BTC forecasting. ARIMA(2,1,2) has achieved the lowest error values,MAE =$30.18 and MAPE = 0.043% on a 15-day test.Another model, Prophet has huge error values with MAE=$8,897.93, as it requires a seasonal structure,which absent from one-minute data. Both LSTM models underperformed ARIMA, with the univariate reaching MAE = $64.52. and multivariate reaching MAE= $68.91.These results are consistent with the efficient market hypothesis: Bitcoin at one-minute resolution behaves as a near-random walk, and no model achieved reliable directional prediction above chance. In the cross-market analysis, NASDAQ daily returns showed the highest correlation with BTC (r = 0.617), but Granger causality tests found no predictive lead-lag relationship. The multi-market daily LSTM reduced forecast error by 6.8% over a BTC-only LSTM baseline, though both remained above the naive persistence baseline on the true out-of-sample daily test.