Lesson 1Deployment considerations: retraining cadence, data pipelines, monitoring drift and backtesting forecastsPlan putting revenue prediction systems to work, including how often to retrain, automatic data flows, watching for changes in data and ideas, and steady checking back to keep good work in shifting market times.
Designing robust data pipelinesScheduling retraining and updatesMonitoring data and concept driftOngoing backtesting and benchmarkingAlerting and rollback strategiesLesson 2Communicating forecasts to stakeholders: visualization of point forecasts and intervals, scenario analysis, transparency of assumptionsShare revenue predictions well with those involved using clear pictures of single points and ranges, what-if checks, and open notes on ideas, limits, and risks from the model to build trust.
Visualizing point and interval forecastsScenario and what-if analysis designExplaining drivers and key featuresDocumenting assumptions and limitsTailoring messages to stakeholdersLesson 3Machine learning time series models: random forests/gradient boosting with lagged features, XGBoost/LightGBM, and sequence models (LSTM/GRU)Teach machine learning models for time-based revenue prediction, using tree groups with past features and sequence models like LSTM and GRU, while dealing with steady changes, seasons, and differences per product.
Random forests with lagged featuresGradient boosting, XGBoost, LightGBMGlobal versus local forecasting modelsSequence models with LSTM and GRUHandling nonstationarity and scalingLesson 4Formulating forecasting objectives and evaluation horizons (e.g., next 3, 6, 12 months)Set prediction aims for main bank services by picking targets, time spans, and details, matching them to business choices like budgeting, pricing, cash planning, and rules or risk reports needed.
Choosing revenue targets and unitsSelecting forecast horizons and frequencyAligning forecasts with business decisionsGranularity by product, segment, and regionHandling new products and short historiesLesson 5Identifying and sourcing time series data (public financial series, payment volumes, synthetic generation techniques)Find out how to spot, check, and get time series data for bank money prediction, including inside product numbers, public money series, and made-up data that safely adds to few or unclear past records.
Cataloging internal product revenue seriesUsing public macro and market data sourcesCollecting payment and transaction volume dataAssessing data quality, gaps, and revisionsSynthetic data generation for stress scenariosLesson 6Training and hyperparameter tuning: grid/random search, Bayesian optimization, time-aware scoringMake model work better using planned tuning of settings, like grid and random search, Bayesian ways, and time-based scoring that keeps order of time and focuses on key business spans.
Defining search spaces and priorsGrid and random search trade-offsBayesian optimization workflowsTime-aware validation and scoringEarly stopping and resource limitsLesson 7Model ensembling and reconciliation: simple model averaging, weighted ensembles, stacking for time seriesMix many prediction models for bank money using simple mixes, weighted groups, and stacking, and use level-based fixing to make sure predictions match across services, branches, and groups in the bank.
Simple and weighted model averagingStacking and meta-learners for seriesDiversity and correlation among modelsHierarchical and grouped reconciliationEvaluating ensemble stability over timeLesson 8Baseline time series methods: ARIMA, ETS, naive and seasonal naïve models, decomposition (trend/seasonality)Look at basic time series models for bank money, like simple, seasonal simple, ARIMA, ETS, and breaking down, to set a standard work level and see trends and seasons before using hard machine learning models.
Naive and seasonal naive benchmarksClassical decomposition of trend and seasonalityARIMA modeling for bank revenue seriesExponential smoothing and ETS variantsComparing baselines across productsLesson 9Feature engineering for revenue: lags, rolling means/std, differencing, calendar effects, holiday indicators, cohort effects, marketing/campaign flagsBuild useful signs for bank money prediction, like past values, moving averages, changes, date and holiday effects, group and life cycle signs, and marketing flags that show demand shifts and big changes.
Lag and lead features for revenueRolling means, volatility, and ratiosCalendar, holiday, and payday effectsCohort and lifecycle based featuresMarketing and campaign impact flagsLesson 10Data splitting and cross-validation for time series: train/validation/test splits, expanding window CV, blocked CVPlan time-smart data splits and checks for revenue prediction, using moving and blocked ways, to stop leaks, copy real use, and get sure guesses of model work over time in bank settings.
Holdout train, validation, and test splitsRolling and expanding window validationBlocked cross-validation for seasonalityPreventing temporal leakage in featuresBacktesting over multiple forecast originsLesson 11Evaluation metrics and error analysis: MAE, RMSE, MAPE, symmetric MAPE, prediction intervals and coverageCheck revenue predictions with measures like MAE, RMSE, MAPE, even MAPE, and range cover, and do deep error checks by group, span, and time to find biases and weak spots in models.
Scale-dependent error metricsPercentage and relative error metricsPrediction intervals and coverageHorizon and segment level diagnosticsRegime and event-driven error analysisLesson 12Macro and external regressors: using CPI, unemployment, interest rates, mobility, Google Trends; feature selection and lag alignmentAdd big economy and outside factors to money models, like CPI, jobless rates, interest, movement, and search trends, and learn ways for past matching, sizing, and picking features to stop overfit.
Selecting relevant macro indicatorsAligning lags between macro and revenueTransforming and scaling external dataFeature selection and regularizationStress and scenario overlays with macros
