weatherman

This is weather predictor for humidity and temperature, simply trained on fbprophet. As a personal learning project in winters 2024. Here are some learnings:

Handles Complex Seasonality: Prophet automatically detects daily, weekly, and yearly seasonal patterns. Incorporates Holidays: Easy to add holidays or special events that affect the forecast. Robust to Missing Data: Prophet handles missing values without requiring special treatment. Automatic Model Tuning: Prophet automatically selects the best parameters for trend and seasonality. Non-Linear Trends: Prophet can model non-linear trends (e.g., exponential growth) easily. User-Friendly: Prophet is easier to use and tune compared to ARIMA, which requires more manual adjustments.

HATING ON ARIMA BECAUSE MORE STATISTICAL AND DOESNT CAPTURE SEASONALITY TYPES. The standard ARIMA model is designed for non-seasonal time series, so if your data has clear seasonal patterns, you'll need to use Seasonal ARIMA (SARIMA) or similar extensions. LSTMs require more computational power and memory than models like Prophet, especially for large datasets. Long Training Times: Training deep learning models can take significantly longer compared to simpler models. Prone to Overfitting: LSTMs can overfit if the dataset is small or if hyperparameters are not tuned correctly.

apmhan and remal -> cyclones in westbengal which contributed to that outlier. just removed it z scores >3;; removed. Z Score = (x − x̅ )/σ Where, x = Standardized random variable. x̅ = Mean. σ = Standard deviation.

changepoint_prior_scale 0.2 is considered high.0.05 is decreased. Controls the flexibility of the model in adapting to trend changes (changepoints). A higher value allows the model to adapt more quickly to changes in the trend. A lower value smoothens the trend and avoids overfitting to noise. seasonality_prior_scale Controls the flexibility of the seasonal components. A higher value allows more variability in the seasonal patterns. A lower value restricts seasonal effects, making them smoother. //Use a higher value (e.g., 20.0) if: Your data exhibits strong seasonal fluctuations. You suspect the default seasonality constraints are too restrictive. //Use a lower value (e.g., 5.0) if: Seasonal patterns are expected to be smooth and not too variable. You want to prevent overfitting to irregularities in the data. *** changepoint_range Defines the proportion of the data where changepoints are considered (default: 0.8). A higher value (e.g., 0.95) allows changepoints closer to the end of the dataset. A lower value focuses changepoints earlier in the data. When to Increase Use a higher value if: You suspect trend changes occur toward the end of your dataset. Your dataset includes recent, significant events (e.g., a pandemic, market crash). When to Decrease Use a lower value if: You believe trend changes are more likely in earlier parts of the data. You want to ignore recent anomalies or noise.