TechGear: Forecasting Sales + Optimizing Marketing Spend (Regression, Simulation, and Linear Programming)

Why I built this

TechGear wanted a clear answer to a very normal business question: what actually drives sales, and how should we allocate budget to get the best return? I treated this like a real analytics engagement—starting with exploration, moving into statistical modeling, testing forecasts, stress-testing uncertainty with simulation, then finishing with an optimization model that converts insight into a decision.

---

Click on the photo to navigate to the slide deck.

Slide 1 — Project overview

This deck is an end-to-end walkthrough of the workflow:

• Explore the data and relationships

• Model sales drivers using regression

• Forecast sales using time-series methods

• Simulate uncertainty with Monte Carlo

• Optimize budget allocation with linear programmingMy goal wasn’t just to “build models.” It was to create a decision-ready story.

Slide 2 — Data and what it includes

Here I define the dataset at a high level (what’s measured, at what cadence, and what the business controls vs. can’t control).This matters because modeling gets messy fast if you don’t clarify:

• what “sales” represents (units vs revenue),

• whether the time period has seasonality,

• and which inputs are marketing levers (ex: ad spend, promotions) vs. external factors.

Key note: before modeling, I checked for missing values, obvious outliers, and whether the variables were comparable (same units, same time scale).

Slide 3 — Initial exploration (relationships and patterns)

This is where I looked for “shape” in the data:

• Do relationships look linear?

• Are there threshold effects (spend only helps after a minimum)?

• Are there outliers that would dominate the model?

Even basic charts here are powerful because they tell you what kinds of models make sense before you start fitting equations.

Insight from this stage: certain inputs show a clearer relationship with sales than others—meaning not all spend is equally effective.

Slide 4 — Simple linear regression (starting point)

I start with a simple regression to establish a baseline relationship: one primary predictor → sales.

This answers: if this lever increases by 1 unit, what’s the expected change in sales?

What I’m looking for on this slide:

• coefficient direction (positive/negative),

• significance / strength,

• and the baseline explanatory power (R²).

Portfolio point: simple models are not “less impressive.” They’re how you check reality before building complexity.

Slide 5 — Diagnostics + assumptions (is this model trustworthy?)

A regression isn’t “good” just because it outputs a line and a coefficient.Here I look at whether the assumptions hold:

• residual patterns (non-linearity)

• variance changes (heteroskedasticity)

• influential points/outliers

If assumptions are violated, I either transform variables, adjust the approach, or move to a more appropriate model.

Slide 6 — Multiple regression (drivers together, not in isolation)

Real businesses don’t pull one lever at a time, so this slide expands into a multi-variable model: multiple predictors → sales.

This is where things get more interesting:

• Some predictors remain strong when controlling for others.

• Some lose impact (their “effect” was actually shared with another variable).

• Multicollinearity can appear (predictors overlapping heavily).

Key takeaway: this model gets us closer to causal-ish interpretation, or at least stronger decision guidance.

Slide 7 — Forecasting approach (and why)

Once you understand drivers, the next question is: what do we expect sales to look like next period?

Here I introduce forecasting methods (like moving averages / exponential smoothing / regression-based forecast, depending on what your slide shows).

What I emphasize:

• the reason the forecast method matches the data structure,

• how trend/seasonality were handled,

• and what horizon is realistic.

Slide 8 — Model comparison (choosing the best forecast)

This slide is about not guessing. I compare forecast models using an error metric (often RMSE/MAE), ideally using a validation approach like holdout or cross-validation.

The point: pick the forecast that performs best out-of-sample—not the one that “looks nicest.”

Decision implication: the chosen model becomes the planning baseline for inventory, staffing, and budget.

Slide 9 — Monte Carlo simulation (stress testing uncertainty)

Forecasts are single-number stories. Businesses live in ranges.

This slide uses Monte Carlo simulation to show:

• likely outcomes,

• downside risk,

• and upside potential.

  
  

Instead of “sales will be X,” we get something like:

• “There’s a ___% chance sales land between A and B”

• “Worst-case scenarios cluster around C”

• “The upside tail suggests…”

This is where decision-makers start to trust the analysis because it matches how uncertainty actually feels.

Slide 10 — Linear programming (turning insight into an allocation decision)

This is the “so what do we do?” slide.

Linear programming lets us optimize a goal (ex: maximize profit or sales) subject to constraints:

• total budget

• minimum/maximum spend by channel

• operational limits

This turns your regression/forecast insight into a concrete recommendation: allocate budget across channels in a way that mathematically maximizes the objective.

Slide 11 — Recommendations and next steps

This is the portfolio-critical finish: clear recommendations tied to business action.

Recommendations (example framing):

1. Prioritize the highest-performing channel(s) identified by the regression + validated forecast performance.

2. Use the forecast range (not a single number) for planning: set base, conservative, and stretch targets.

3. Operationalize the optimization model as a monthly/quarterly planning tool (re-run with updated data).

4. Monitor leading indicators (the variables that move first before sales does).

5. Refresh models on a schedule (quarterly is a strong default) to prevent drift.

Next steps:

• add new periods of data,

• test non-linear models if diagnostics suggest it,

• and refine constraints so the LP model matches real-world spending rules.

---

Deliverables (what I would hand to a stakeholder)

• Slide deck (this presentation)

• Clean dataset + documentation (data dictionary)

• Model outputs + assumptions notes

• Forecast comparison summary

• Optimization constraints + recommended allocation