How to Build a Moving Average Model for Predicting NFL Passing Yards

Introduction

Quarterback passing yards props are among the most popular bets in NFL markets, with millions wagered every week on FanDuel, DraftKings, and other sportsbooks. The appeal is obvious: unlike team outcomes that depend on 53 players, special teams, and coaching decisions, passing yards feel more predictable—it's primarily one player's performance. But here's the reality check: even the best statistical models struggle to consistently beat the oddsmakers.

In this tutorial, I'll walk you through building a weighted moving average model to predict QB passing yards using Parlay Savant, a tool that removes the technical friction of data collection and analysis. We'll pull historical NFL data, construct a simple but theoretically sound prediction model, test it against actual Week 5 games from the 2025 season, and honestly assess whether it could beat Vegas lines.

By the end, you'll understand how to build your own passing yards predictor, interpret its results, and—most importantly—recognize its limitations. Fair warning: sports prediction is brutally difficult. Even a "good" model might only hit 55-60% accuracy on over/under bets, and we'll see whether ours reaches that threshold. The goal here isn't to promise you a money-printing machine, but to teach you the process of building, testing, and critically evaluating predictive models.

Step 1: Getting the Data with Parlay Savant

The first challenge in any sports analytics project is data collection. I needed comprehensive QB performance data from recent games to build a model based on recent form. Here's the exact prompt I used:

Get QB passing performance data from the 2025 season weeks 1-4, including passing yards, attempts, completions, touchdowns, and opponent for each game. Filter for QBs with at least 10 attempts per game. Include game recency ranking so I can weight recent games more heavily.

Parlay Savant processed this request by generating a SQL query against its NFL database. The query joined multiple tables—

player_game_stats

players

games

teams

ROW_NUMBER() OVER

The resulting dataset contained 29 qualifying quarterbacks with performance data across 4 weeks of the 2025 season. Here's a sample of what came back:

One thing that immediately surprised me: the massive week-to-week volatility. Russell Wilson had a standard deviation of 165.2 yards across just three games—ranging from 160 to 450 yards. Even elite QBs like Josh Allen showed 106-yard standard deviations. This volatility is exactly why prediction is so difficult.

The visualization above shows just how unpredictable QB performance can be. The top chart displays the distribution of passing yards for the five most active QBs—notice the wide spreads. The bottom chart tracks week-by-week trends, and you can see wild swings: Dak Prescott went from 188 yards in Week 1 to 361 in Week 2, then back down to 251 in Week 3. These aren't random backups; these are starting QBs with consistent playing time.

Key data insights:

• 116 total QB performances across 4 weeks
• Average passing yards: 232.2 yards (median: 228.5)
• Range: 67 to 450 yards (that 450 was Russell Wilson's outlier game)
• QBs with highest volatility were often in unstable offensive situations or facing dramatically different defensive matchups

This volatility would become a major challenge for my model. A simple moving average assumes some consistency in performance, but NFL games are influenced by game script, weather, opponent strength, and dozens of other factors that change weekly.

Step 2: Building the Model with Parlay Savant

With the data in hand, I needed to design a prediction approach. I chose a weighted moving average because it's conceptually simple but incorporates an important insight: recent performance matters more than older games. A QB coming off a 350-yard game is probably in better form than one who threw for 350 yards three weeks ago.

Here's the prompt I used to build the model:

Build a weighted moving average model to predict Week 5 passing yards. Use these weights: most recent game 40%, second most recent 30%, third most recent 20%, fourth most recent 10%. Require minimum 3 games to make a prediction. Calculate predictions for all qualifying QBs and show the top 15.

Parlay Savant generated Python code that implemented this logic. Here's the core calculation it created:

weights = {1: 0.40, 2: 0.30, 3: 0.20, 4: 0.10}

for qb in df_historical['name'].unique():
    qb_games = df_historical[df_historical['name'] == qb].copy()
    
    if len(qb_games) >= 3:
        weighted_sum = 0
        total_weight = 0
        
        for _, game in qb_games.iterrows():
            recency = int(game['game_recency'])
            if recency in weights:
                weighted_sum += game['passing_yards'] * weights[recency]
                total_weight += weights[recency]
        
        predicted_yards = weighted_sum / total_weight

The model iterates through each QB's recent games, multiplies each performance by its weight, and calculates a weighted average. It's not machine learning—there's no training phase or optimization. It's pure statistical averaging with recency bias.

Technical metrics from the model build:

• 29 QBs generated predictions (all with 3+ games)
• Average prediction: 216.9 yards
• Prediction range: 155.1 to 293.9 yards
• Model complexity: O(n) time complexity, no parameters to tune

Let me show you exactly how the calculation worked for the top-predicted QB, Dak Prescott:

Dak Prescott Weight Calculation:

• Week 4: 319 yards × 40% = 127.6
• Week 3: 251 yards × 30% = 75.3
• Week 2: 361 yards × 20% = 72.2
• Week 1: 188 yards × 10% = 18.8
• Total: 293.9 predicted yards for Week 5

Notice how the model heavily weights his most recent 319-yard performance while still incorporating the earlier games. This makes intuitive sense—Prescott was clearly in good form heading into Week 5.

The model also predicted Matthew Stafford at 292.9 yards (coming off 375, 289, and 196 in his last three games) and Jordan Love at 266.9 yards. These felt reasonable based on recent trends.

What I liked about this approach: transparency. I could explain exactly why each prediction was made. There's no black-box algorithm. What concerned me: the model has no awareness of opponent strength, home/away splits, weather, or game script. It's purely backward-looking at the QB's own performance.

Step 3: Making Predictions

With the model built, it was time to generate actual predictions for Week 5 games. I wanted to see not just the numbers, but how they compared to Vegas betting lines. Here's my prompt:

Generate Week 5 passing yards predictions for all QBs using the weighted moving average model. Create a visualization showing the top 4 predicted QBs with their historical trends and the Week 5 prediction overlaid. Also show how predictions compare to season averages.

Parlay Savant created a multi-panel visualization showing historical performance with the Week 5 prediction marked as a star:

This graph was incredibly revealing. Look at Dak Prescott's chart (top left): the model predicted 293.9 yards based on his recent upward trend, but his season average was only 279.8. The model was essentially betting on continued hot form.

Matthew Stafford's prediction (292.9 yards) looked aggressive given his volatility—he'd ranged from 196 to 375 yards in four games. Jordan Love and Daniel Jones showed more stable trends, making their predictions feel more reliable.

Top 10 Week 5 Predictions:

Looking at these predictions, I had immediate concerns about several:

Dak Prescott (293.9): The FanDuel line was 241.5 yards. My model was predicting 52 yards OVER the Vegas line. That's a massive discrepancy. Either I'd found real value, or I was missing something Vegas knew (spoiler: it was the latter).

Justin Herbert (251.4): His prediction seemed reasonable, but his high standard deviation (52.9 yards) meant the confidence interval was wide. He could easily land anywhere from 200-300 yards.

Russell Wilson (258.4): That 165-yard standard deviation terrified me. Predicting him felt like throwing darts blindfolded.

The model generated predictions for 29 QBs total, but I was most interested in testing it against the QBs with active betting markets—the ones where FanDuel had posted over/under lines.

Step 4: Testing Predictions

The moment of truth. Week 5 games were completed, and I could now compare my predictions to actual results. Here's the validation prompt:

Compare my Week 5 predictions to actual passing yard results. Calculate accuracy metrics including MAE, RMSE, and percentage error. Show which predictions were closest and which were biggest misses. Also compare my predictions to FanDuel betting lines to see if the model would have made profitable bets.

Parlay Savant pulled the actual Week 5 results and merged them with my predictions. The results were... humbling.

Model Performance Metrics:

• Mean Absolute Error (MAE): 66.1 yards
• Root Mean Squared Error (RMSE): 76.7 yards
• Mean Absolute Percentage Error: 24.9%
• Predictions within 20 yards: 1 of 20 (5.0%)
• Predictions within 50 yards: 9 of 20 (45.0%)

To put this in perspective: an MAE of 66 yards means my average prediction was off by about 28% of the actual value. That's not terrible for a simple model, but it's not good enough to consistently beat Vegas.

Success Stories

Bo Nix was my best prediction: I predicted 235 yards, he threw for 242—only 7 yards off. This worked because Nix had been remarkably consistent (215, 138, 246, 246 in his four games), and the weighted average captured his recent upward trend.

Bryce Young (predicted 177, actual 198) and Geno Smith (predicted 206, actual 228) were also solid calls. These QBs had relatively stable performance patterns, which suited the moving average approach.

C.J. Stroud (predicted 215, actual 244) was close enough to feel good about—within 29 yards despite his game-to-game variance.

Epic Misses

Baker Mayfield: I predicted 245 yards. He threw for 379 yards—off by 134 yards (35% error). This was a complete whiff. Looking back, Mayfield was facing Seattle's pass-funnel defense in a game that turned into a shootout. My model had no way to account for game script or opponent-specific matchups.

Justin Fields: Predicted 158, actual 283 (off by 125 yards). Fields had been inconsistent all season, and my model underweighted his ceiling. He also benefited from a favorable matchup against Dallas.

Jalen Hurts: Predicted 155, actual 280 (off by 125 yards). This miss hurt because Hurts is a running QB whose passing volume is highly game-script dependent. When the Eagles get ahead, he throws less. When they're behind, he airs it out. A simple moving average can't capture this.

Cameron Ward and Sam Darnold were also major misses, both underestimated by 100+ yards.

Betting Performance Analysis

The real test: would this model make money betting against Vegas lines?

Betting Results:

• Correct Over/Under calls: 9 of 17 (52.9%)
• Breakeven threshold: 52.4% (at standard -110 odds)
• Verdict: Barely profitable, but not by enough to overcome variance

Here's the detailed breakdown:

Winning bets: The model correctly identified several QBs who would exceed their lines, particularly Stafford, Mayfield, and Darnold. These were all cases where recent form suggested upward momentum.

Losing bets: The Dak Prescott loss stung the most. I predicted 293.9 yards (betting OVER 241.5), but he only threw for 237. This was a case where the model was overly influenced by his recent hot streak and didn't account for opponent adjustments or regression to the mean.

The cluster of UNDER bets that lost (Goff, Tua, Allen, Browning) revealed a systematic problem: my model was too conservative on QBs with recent down games. When these QBs bounced back, I was on the wrong side.

Model vs. Vegas Accuracy

FanDuel Line MAE: 46.3 yards
My Model MAE: 56.6 yards
Difference: +10.3 yards (Vegas was more accurate)

This is the brutal reality: Vegas lines beat my model by about 10 yards per prediction on average. That might not sound like much, but in betting terms, it's the difference between long-term profit and loss.

Why was Vegas better? They incorporate everything I didn't: opponent pass defense rankings, home/away splits, weather forecasts, injury reports, line movement from sharp bettors, and proprietary models built by teams of PhDs. My weighted moving average only looked at the QB's recent box scores.

That said, 52.9% accuracy on over/under calls is actually not bad for a simple model. It's above the breakeven threshold, which means in theory, you could make a tiny profit over a large sample. But the edge is so small that variance would dominate in the short term.

Conclusion: What We Learned

After building, testing, and validating this weighted moving average model for NFL passing yards, here's my honest assessment:

What worked:

• The model achieved 52.9% accuracy on over/under bets, barely above the 52.4% breakeven threshold
• Simple, transparent methodology that anyone can understand and replicate
• Captured recent form effectively for stable QBs like Bo Nix and Geno Smith
• Parlay Savant removed all the technical friction—I spent zero time writing SQL or debugging Python

What didn't work:

• MAE of 66.1 yards is too high for consistent betting success
• No awareness of context: opponent strength, game script, weather, injuries
• Overweighted recent outliers: Dak Prescott's hot streak led to a bad prediction
• Underestimated volatility: QBs like Baker Mayfield and Justin Fields have huge variance that moving averages can't capture
• Vegas lines were 10 yards more accurate on average

The harsh reality: About 75% of the variance in QB passing yards remains unexplained by this model. Week-to-week performance is influenced by dozens of factors beyond recent form—defensive matchups, weather conditions, game flow, offensive line health, receiver availability, and pure randomness.

Could this model be improved? Absolutely. Here's what I'd try next:

1. Add opponent adjustments: Weight predictions based on opponent pass defense rankings
2. Incorporate home/away splits: QBs perform differently on the road
3. Include weather data: Wind and rain significantly impact passing games
4. Use exponential smoothing: More sophisticated than linear weights
5. Add confidence intervals: Acknowledge uncertainty in predictions
6. Test different weight schemes: Maybe 50/30/20 works better than 40/30/20/10

But even with all these improvements, I'm skeptical we'd consistently beat Vegas. The sportsbooks have massive advantages in data, modeling sophistication, and market information.

The value of Parlay Savant: What impressed me most wasn't the model's accuracy—it was how quickly I could iterate. From initial idea to tested predictions took maybe 30 minutes of actual work. Normally, this would require hours of data wrangling, SQL queries, and debugging. Parlay Savant handled all that, letting me focus on the modeling logic and interpretation.

Final verdict: This weighted moving average model is a solid learning tool and a reasonable starting point for understanding QB prediction. At 52.9% accuracy, it's technically profitable, but the edge is too thin to recommend betting real money on it. You'd need a much larger sample size to overcome variance, and even then, you're fighting against Vegas's superior information.

If you're building this for educational purposes or as a foundation for more complex models, it's valuable. If you're hoping to get rich betting QB props, keep your day job. Sports prediction is humbling work, and even simple models reveal just how much we don't know about what will happen on Sunday.