Sales Forecasting Engine

01 — Problem

Product Context

The business relied on manual sales forecasts based on historical averages and intuition. These forecasts had a Mean Absolute Percentage Error (MAPE) of 14.1%, leading to chronic issues:

Over-forecasting caused excess inventory and wasted resources. Under-forecasting led to stockouts and missed revenue. The operations team needed reliable, automated forecasts across multiple product lines and regions to enable proactive planning rather than reactive adjustments.

02 — System Architecture

Data & Pipeline Design

Designed a modular forecasting pipeline that ingests historical sales data, applies preprocessing and feature extraction, and produces forecasts at daily/weekly/monthly granularity.

Sales Database
PostgreSQL

→

ETL Pipeline
Cleaning + aggregation

→

Feature Extraction
Temporal features

→

ARIMA / Prophet
Ensemble forecast

→

Forecast Output
Time-series predictions

→

Dashboard
Looker Studio

03 — Modeling

Modeling Approach

Evaluated multiple time-series models and built an ensemble approach combining statistical methods (ARIMA/SARIMA) with Facebook Prophet for capturing trend, seasonality, and holiday effects.

Models ARIMA/SARIMA for stationary series, Prophet for series with strong seasonality and holiday effects, weighted ensemble for final forecast

Challenges Multiple seasonalities (weekly, monthly, quarterly), holiday spikes, trend changes during promotions, missing data gaps

Features Day-of-week, month, quarter, lagged values (7/14/30 day), rolling averages, holiday indicators, promotion flags

Optimization Minimized MAPE across forecast horizons (7-day, 30-day), with backtesting on 12-month rolling window

Validation Walk-forward validation, residual analysis, prediction interval calibration (80% and 95%)

04 — Output Layer

Forecast Dashboard

The forecast outputs are visualized in a Looker Studio dashboard showing actual vs. predicted trends, forecast confidence intervals, and error tracking by product line and region.

Current MAPE

8.3%

Previous MAPE

14.1%

Improvement

41%

Actual vs. Forecast (30-Day Window)

Forecast

Actual

80% CI

05 — Decision Dashboard

Interactive Dashboard

This dashboard provides a clear view of revenue, costs, and profit trends over time. It supports financial decision-making by highlighting performance changes, category breakdowns, and geographic distribution.

View Interactive Dashboard →

06 — Insights

Key Findings

Weekly seasonality was the strongest signal — Monday/Tuesday consistently under-indexed while Thursday/Friday peaked
Prophet's holiday component captured promotion-driven spikes that ARIMA alone couldn't model
The ensemble approach reduced error most in volatile product lines where single models diverged significantly
Forecast accuracy degraded beyond 14 days — operational planning was redesigned around 7-day rolling forecasts

07 — Decision Layer

How forecasts drove action

The forecasting engine became the input layer for operations planning. Inventory was pre-positioned based on 7-day rolling forecasts. Staff scheduling aligned with predicted demand peaks. Marketing promotions were timed around forecast valleys to smooth demand curves.

The key shift: from "what happened last month" to "what will happen this week" as the basis for resource decisions.

08 — Impact

Business Impact

MAPE Reduction

14.1% → 8.3%

41% improvement in accuracy

Planning Horizon

7-day

Rolling automated forecasts

Overstock Reduction

↓ 23%

Less wasted inventory

09 — Learnings

Reflections

"The most impactful decision wasn't choosing between ARIMA and Prophet — it was deciding the forecast horizon. We initially built for 30-day forecasts, but accuracy dropped sharply past 14 days. By shifting to 7-day rolling forecasts aligned with the business cadence, we delivered dramatically more useful predictions. The lesson: match the model's capability to the decision it supports."

Sales Forecasting Engine(Time-Series ML)