October 2, 2025

Time to reflect on the predictability of the 2025 season so far!

If you’re new to these: Examining predictive accuracy has been a long-standing tradition, underpinning a key purpose of Subvertadown: To give us an understanding of how predictable things have been. We want to know "Was this season more or less predictable than normal?"

Compared to Other Seasons

So as usual, here’s a look at how each individual model is doing, compared to other seasons.

This does not tell "how good the models are". It only tells us how predictable the current season is, compared to the historical norm.

We’re seeing a different pattern this particular year. This time, the yellow bars do NOT all closely match the expected level at the blue bars.  Some positions are not behaving with the same predictabilities as usual.

This is an interesting mix—it’s not like last year when I reported “everything normal here”.  QBs, WR, and RB are the positions closest to average, essentially quite in line with what we could hope to expect the first 4 weeks.   We’re seeing especially high predictability for D/STs, which must be related to the very high predictability of Vegas lines (bettors being able to predict game scores).  However, falling behind in accuracy are: Kickers and TE.  The Kicker result is due to week 3-- a single very-backwards week.  And likewise, TE is affected by a single negative week (week 4).   Most notable is how much my own Game Scores models have even been lucky enough to outperform Vegas, for the season start.  This is an unusually high accuracy, and it has an influence on the other positions-- Obviously it had a large positive effect on betting lines.

Comparative Accuracy Assessment

Reminder / for newbies: I’m not trying — and don’t expect to be— “#1”. That’s plain unrealistic to expect that all the time, against very competent expert analysts I track. My goal is rather to check that the models are still performing at a similar level to others, specifically sources that have been consistently good for at least a few years. Many sources are great one year but then poor the next. My chosen experts are good in a more consistent way. We naturally trade places at the top, year to year. Knowing that the models perform at least at a "similarly" to top sources lends confidence and gives reason to trust forecasting, when we extrapolate models to future weeks.

Defensive Maneuvers

Last year 2024 was also good, and now in 2025 we can say we’re lucky at streaming D/ST for 2 years in a row.  D/ST accuracy has been amazingly excellent to start this year. And that goes for all rankers, as well as myself!  My models have had a fantastic showing, and it’s surprising how clear it emerges from the accuracy measurements, showing my models standing at #1 overall.    If you’ve followed me, you know I don’t always get to say that!  At this point last year, I was #4 in my standings.  But anyway, that’s not the main point. My conclusion is only that the D/ST models are performing, and they’re not in need of any kind of update.

Note that these charts are for the standard “Yahoo” (or Sleeper) settings.  I am not displaying the ESPN setting (although my model leads for that, as it had hopefully better!). 

Here’s the Kicker

Weeks 1 and 2 were looking on fire to start the season!  Then a real problem came in week 3.  Almost all rankers saw a negative correlation (except one), which meant accuracy was severely hit. But it was my model in particular that had the most negative correlation for that particular week.  As a result, my overall season accuracy is exactly average for the season, compared to the other sources I track.   

Here is the plot I usually show for Kickers, which is the “Accuracy Gap” method invented for FantasyPros.  I’ve had more ups and downs, and it looks inconclusive.

One key purpose of these posts has historically been to reflect on areas for improvement.  I’m not exactly worried about the kicker models, but I do have ideas for what could have been better.  My best guess is that the kicker utilization is tied to the predictability trend of Vegas.  And as pointed out above, Vegas has been more accurate than usual, this year.  Last year, Vegas lines started out a few weeks with very poor predictability.  To me, high Vegas accuracy feels like an indicator of whether we’re “past the early season” phase, and therefore my dedicated “early season kicker model” would not do its intended job.  (The early-season model is meant to deal with uncertainty in team strengths.)  And, in fact, now I can report this appears to be true.  I can see that, if I had instead applied my dedicated “mid-season kicker model”, then Kicker accuracy would have been much better this year. So, in hindsight, I could have reacted to the high accuracy of Vegas during weeks 1 and 2, and then assumed week 3 was already like “mid-season”. 

Two Cents for a Quarterback

For yet another year, my QB model is booming.  It seems like it’s always kind of a pleasant surprise when I compare it against other top QB experts, since I don’t actually set my expectations to be #1… and yet something seems to be cooking with the QB model each year. And 2025 might have been the best year that I’ve seen start off, as I’ve maintained #1 performance.  If I could choose, I would prefer that my Kicker model show itself so clearly as this!  But after years of reviewing this, I’m comfortable saying that: apparently my methodology suits QB really well.

The first chart here shows correlation coefficients, and it shows relative overperformance weeks 2-4 against the average. The right-hand chart shows the Accuracy Gap method, where I’ve included 4 of the other sources since it didn’t get too cluttered.

Survivor

Based on probability alone, we should normally expect half the pathways to lose once, during the last 4 weeks of Survivor.  Two or three losses total would be a normal amount, and it would mean that I’d need to display my “backup” pathways that I show each week (1b, 2b, 3b). But amazingly, the actual number of losses has yielded just 1 single replacement!  That was for pathway #3-- when Vikings lost week 2.  Pathways 1 and 2 have remained unchanged, so far. The current results are better than normal.

Normally I show losses (there were 7 last year), but for this period of time, I can just list the pathways:

• Pathway 1: Broncos, Ravens, Bills, Lions

• Pathway 2: Commanders, Lions, Buccaneers, Bills

• Pathway 3: Eagles, 49ers (replaced Vikings), Seahawks, Texans

Betting Lines

Our baseline expectations should normally be to lose about -20% of the weekly pot, by this time (after 4 weeks). That means, a normal “coin flip” betting process would lose us about -5% per week, from the target weekly bet amount. Yes, that’s if we were just monkeys shooting darts.

This has been an exceptionally lucky year, so far, with an average ROI of +20% per week (of the weekly post).  This worries me for future betting behavior of my subscribers, because it cannot statistically be sustainable all season, and we should rather expect regression.  In fact, most of the gains were from week 1 (otherwise the weekly avg would be 10%).

Recorded Bet Picks and W/L outcome:

Extra Commentary on Betting Lines Updates

Since it’s been going so well, I thought it was worth explaining what I changed since last year, to avoid the negative returns that it saw in 2024 from a very bad 49% win rate.  The year 2024 should have done at least as well as 2023 was!  So I made updates based on the examinations I published about last year.  Here, and here.  I especially caught my own clue from the first link, where I realized that it would be hard to do worse even if I tried— Then I thought, maybe my model IS actually trying!

The easiest explanation is to say that last year it was “too cute”: I had things set up to “bet against myself”, in a logic that was meant to hit bigger in a positive way.  I know that must sound strange, but it means I had 2 models running against each other, because one was supposed to be a “dumber” model that predicted how “worse” bettors would behave.  Theoretically, this would improve returns, but it turned out that “dummy” model did too well—it was actually more accurate than Vegas itself!  That’s disappointing because it shows the opportunity was actually there— and bad news for something you’re “subtracting”. So I feel quite sure this caused my models to defeat themselves in 3 or so specific weeks where we saw a “crash” of what we’d earned so far.

So, in short, I got rid of this “too cute” logical setup, by deleting the dummy model effect, while also better honing the main game score models. I won’t go into the extra tailoring here, but it was the most comprehensive I’ve ever been.  I believe this avoids the risk of the 2024 situation repeating.  And I suspect this is the main reason we haven’t had a negative week yet, although obviously that could happen at any time—and everyone needs to remember that!

New Design of Website Table for Betting Lines

As a special message, I want to highlight that I have changed the webpage format for dealing with Open bets (bets recommended for Tuesdays).  The purpose is to better interpret the recommendations.   For years, I’ve used a consistent way of grading accuracy like I did here today, and I wanted that method to be obvious and transparent.  For example, when betting lines change after Tuesday, sometimes the recommendation can be to buy more of a certain pick—and sometimes the recommendation is to sell back the pick, when the lines have moved the wrong direction compared to what we expected.  I hope the new table makes it clear exactly how I gather the numbers used to report ROI.

Looking forward to the next 4 weeks!

/Subvertadown