Our first postseason modeling post focused on the output.

We rebuilt our October pitching model, ran the postseason 100,000 times, and compared the resulting probabilities with the DraftKings postseason board.

But there's another side of this project that might be even more interesting:

Where do the inputs come from?

In the regular season, projecting a pitching staff is relatively straightforward.

You have a scheduled starter. You have a reasonable expectation for how deep he'll work. You have a bullpen.

October changes that.

Managers shorten rotations. A starter might become a reliever. An ace might come back on short rest. A pitcher expected to provide bulk innings might never enter the game. And someone whose workload appears limited might suddenly give his team five innings.

All of that matters to the model.

Rating the Pitcher Isn't Enough

We've spent years developing our Basewinner pitching ratings.

But knowing how good a pitcher is only solves part of the problem.

We also need to know:

How many innings is he likely to pitch?

That's where sxIP — starting expected innings pitched — becomes extremely important.

Suppose we have an elite starter backed by an average bullpen.

If we project him for five innings instead of four, we've effectively replaced one inning of average bullpen pitching with one inning from an elite starter.

That changes the team's expected runs allowed.

Now multiply that across five games in a Division Series.

Then across every possible series in the bracket.

Then run the postseason 100,000 times.

A seemingly small assumption can work its way through the entire simulation.

October Makes sxIP Difficult

Regular-season usage gives us a starting point.

But simply using regular-season averages in October doesn't make much sense.

Teams with first-round byes have rested starters.

Elite pitchers can get longer leashes.

Struggling starters can get much shorter ones.

And the availability of the bullpen can completely change the plan.

So for the postseason we rebuilt the sxIP component.

Our current process begins with the individual pitcher's bwERA+.

We then estimate his postseason workload and combine those innings with the remaining innings at the team's Basewinner bullpen rating.

Conceptually:

Starter bwERA+ × expected postseason innings

plus

Bullpen rating × remaining innings

That gives us a projected pitching-staff rating for that game.

We repeat the process across the expected rotation to build a team-level postseason runs-allowed projection.

Then October Immediately Tests Your Assumptions

Jesús Luzardo was a good example.

There was legitimate uncertainty about how much Philadelphia could expect from him entering the postseason.

Then he went out and gave them five innings on 77 pitches.

That's enormously useful information for the next projection.

Not simply because of how well or poorly he pitched.

We learned something about how Philadelphia was willing to use him.

That's a different piece of information.

Erick Fedde Showed the Other Side

Chicago presented another interesting problem with Erick Fedde.

Based on the available information and his recent role, he looked like a logical candidate to provide bulk innings.

Then he wasn't used in Game 1.

The next game?

He recorded the final six outs.

That's October.

And from a modeling standpoint, it matters.

If we expected two or three innings from one pitcher and the manager uses him completely differently, our bullpen distribution changes.

The answer isn't to pretend we should have known exactly what the manager would do.

The answer is to incorporate the new information.

That's Why the Model Has to Keep Moving

The postseason model isn't something we build once and leave alone for a month.

Every game gives us information.

Who started?

Who was available?

Who wasn't used?

How long was the starter allowed to go?

Who entered when the starter left?

Was a traditional starter used in relief?

How many pitches did everyone throw?

And what does that imply about tomorrow?

Those observations can change sxIP, bullpen availability and ultimately our team pitching projections.

Cleveland Is a Perfect Example of the Problem

Even projecting a five-game Division Series rotation can become complicated.

We may have confidence in Games 1 and 2.

Game 3 can be considerably less certain.

Does the team use another traditional starter?

Does it combine multiple pitchers?

Does somebody who normally starts provide three innings?

Does an elite starter come back earlier than expected?

There isn't always one correct answer available before the series starts.

So instead of pretending that uncertainty doesn't exist, we're trying to model it intelligently.

Sometimes that means creating a blended expectation rather than assigning six innings to a pitcher who may never get six.

Then It All Flows Into the Simulation

This is where the entire Basewinner postseason pipeline starts coming together:

Individual pitcher performance
↓
bwERA+
↓
postseason sxIP
↓
bullpen exposure
↓
game-level pitching rating
↓
team runs allowed
↓
game probability
↓
series probability
↓
100,000 postseason simulations

And from those simulations we can price:

Division Series winners

League champions

World Series champions

Double-chance World Series markets

and even

exact LCS winner/loser combinations.

That's the part of this project I've really enjoyed.

We're not simply asking the model, Who wins the World Series?

We're building a system where changing our expectation for one pitcher's workload can appropriately flow through the rest of the postseason.

The Numbers Will Keep Changing

And they should.

If tomorrow gives us new information about how a team intends to deploy its pitching staff, we'll incorporate it.

If a starter gets pushed deeper than expected, we'll adjust.

If a supposed bulk pitcher is being used as a traditional reliever, we'll adjust.

If a team changes its rotation, we'll adjust.

That's not a weakness in the model.

That's the point of building the pipeline.

The goal isn't to predict every managerial decision perfectly before it happens.

The goal is to take the best information we have, quantify it, simulate what it means—and then be ready to update the model when October gives us something new.