Can You Beat a Bot That Copies FP3 Results? I Tested 6 Prediction Strategies Over a Full F1 Season

Here's a question that might keep you up at night: what if the best prediction strategy is no strategy at all?

I ran an experiment. I created six bots, each following a dead-simple mechanical rule for predicting every qualifying and race session across the entire 2025 F1 season. No gut feelings. No "I think Hamilton's going to have a good weekend" energy. Just raw data copied straight into predictions.

Two of those bots beat every single human in my friend group. And it wasn't a fluke. It was 24 race weekends, 48 sessions, thousands of individual position calls.

So do your predictions actually matter? Let's find out.

The experiment

My friend group has been running F1 predictions for years. Seven of us making qualifying and race picks for every round. The usual setup: watching practice sessions, reading the data, arguing about who's going to nail qualifying in the group chat. When I built Podium Prophets, I imported all the 2025 predictions into the app so I could score them properly.

To be fair, last year none of us did much real analysis. We'd glance at practice results, factor in general knowledge of which teams are strong, and go with gut feeling. No deep-dives into long-run pace, no sector time comparisons, no tyre degradation analysis. Just vibes and a rough sense of the grid.

This year with Podium Prophets it's different. The app has built-in session analysis, race pace charts, and qualifying breakdowns. The question is whether having proper analytical tools actually gives you an edge over just blindly copying the data.

After one of my friends admitted he sometimes just copies the practice results when he's short on time, I got curious. How much do predictions actually improve on raw session data? So I created six bot accounts. Each one follows a single mechanical rule, with zero human judgment involved. I fed their predictions through the exact same scoring system as everyone else: 5 points for an exact position match, 3 for one-off, 1 for two-off, across P1 through P10.

Then I scored the entire 2025 season and compared everyone's totals. The league used carry-forward scoring with accumulating penalties for missed sessions — one of several configurable scoring options in the app. This better reflects real human behaviour: people get busy, forget a weekend, or submit late. The exact point totals would differ under different settings, but the relative patterns hold.

The six strategies

Each bot represents a different "lazy prediction" philosophy. Some use weekend-specific data, some don't. The question is which data sources actually carry predictive value.

A few notes. Standings Stan and Last Race Larry have no data for Round 1, so they skip it entirely and eat the zero. Sprint weekends have a different session structure than normal weekends. If you're unfamiliar with the format, the sprint weekends guide breaks down how they differ.

The sprint-specific bots are where it gets interesting. On sprint weekends, both Lazy Lando and Sprint Sage swap their data sources. Instead of using practice results, they use actual competitive session results from the sprint. The idea: sprint qualifying and sprint races are real competitive sessions, not practice. They should be better predictors.

The results

Here is the full 2025 season leaderboard. Bots are in bold.

Sprint Sage finished first. Lazy Lando finished second. Both ahead of all seven humans.

Let that sink in for a moment.

What the data actually tells us

Before you close the app and hand your predictions to a spreadsheet, let's break down what's actually happening here. The story is more nuanced than "bots win, humans lose."

Later data is dramatically better than earlier data

The clearest signal in the entire experiment. Look at the gap between bots that use different practice sessions:

• FP1 Freddie (Friday morning data): 455 points
• Lazy Lando (FP3 data + grid): 914 points

That's a 459-point gap. FP1 Freddie finished dead last among bots. Lazy Lando finished second overall. Same mechanical approach, different data timing.

This makes sense if you think about it. FP1 is exploratory. Teams run aero rakes, test different setups, split programs between drivers. By FP3, teams have converged on their race weekend configuration. The classification is much closer to real qualifying pace.

If you want the full breakdown of what each practice session tells you, the practice data reading guide covers this in detail.

The starting grid is a remarkably good race predictor

Look at the race points column. Quali Quinn, who literally just copies the qualifying results as their race prediction, scored 549 race points. That's the highest race score of anyone, bot or human. Higher than Sprint Sage's 539, higher than Lazy Lando's 536.

Think about what that means. A bot that predicts zero overtakes, zero strategy swings, zero safety car chaos scored higher than every human who tried to account for those variables.

Overtakes obviously happen. But across a full season, the grid is right more often than your adjustments improve it. When you move Verstappen from P4 to P1 because "he always comes through," you're as likely to lose points on positions 2-4 as you are to gain them on position 1.

Sprint sessions are the best crystal ball

Sprint Sage beat Lazy Lando by 6 points. Small margin, but consistent. The difference comes entirely from sprint weekends, where Sprint Sage uses actual competitive session results instead of practice data.

On normal weekends, Sprint Sage and Lazy Lando use identical strategies (FP3 for qualifying, grid for race). On the six 2025 sprint weekends (China, Miami, Belgium, USA, Brazil, Qatar), Sprint Sage swaps to sprint qualifying results for the qualifying prediction and sprint race results for the race prediction.

This makes intuitive sense. Sprint qualifying is a real qualifying session under pressure. Sprint races produce a real finishing order with real racing. Practice sessions, even FP3, are still practice.

If you're predicting on a sprint weekend, pay more attention to the sprint results than to practice pace. The sprint weekends guide explains why the format creates better data for predictions.

Championship standings alone are not enough

Standings Stan scored 682 points. Not terrible. But firmly mid-table, behind four humans and three bots. "Just pick the drivers in championship order" sounds reasonable, but it misses weekend-specific form entirely.

A driver's season-long points tally tells you they're generally fast. It doesn't tell you whether this specific circuit suits their car, whether they nailed setup, or whether they're carrying a grid penalty. Weekend-specific data matters.

Previous race results are poor predictors

Last Race Larry finished 11th with 586 points. The assumption that "last race predicts next race" doesn't hold up. F1 performance is circuit-dependent. A car that dominates Monza's low-downforce layout can struggle at Singapore's tight corners the very next weekend.

Momentum is real in F1, but it's car-and-circuit specific, not a general trend you can ride from one weekend to the next.

Humans are better at qualifying predictions

Here's the plot twist. Look at the qualifying points column. The top qualifying scorer is human:

• Tóth Bence: 413 qualifying points
• Kenyó Csaba: 399 qualifying points
• Sprint Sage: 381 qualifying points
• Lazy Lando: 378 qualifying points

Humans beat the bots at qualifying predictions by 30+ points. Where the bots dominate is race predictions. Sprint Sage scored 539 race points, Lazy Lando 536. The best human race score was 499 (Kenyó Csaba).

This reveals something important about how humans make prediction errors. People are decent at reading practice pace and translating it to qualifying performance. Where it falls apart is predicting race outcomes. Humans overfit to narratives, overweight dramatic scenarios, and move too many drivers around from their grid positions.

So are predictions pointless?

No. And here's why.

The margin between Sprint Sage (920) and the best human (906) is just 14 points over 48 sessions. That's less than 0.3 points per session. One additional one-off hit every four races would close the gap entirely.

The bots don't have bad weekends. But they also can't capitalize on good reads. They can't see that it's going to rain in Q3. They can't notice that a team found something overnight. They can't factor in a penalty that hasn't been announced yet.

Human prediction isn't worse than mechanical prediction. Human prediction is noisier. Higher highs (beating bots on qualifying) and lower lows (giving back those gains on race day). The path to beating a bot isn't to predict more aggressively. It's to predict more consistently.

How to actually use this

If this experiment changes one thing about how you approach predictions, it should be this: start with the data baseline, then adjust with your knowledge.

Here's a practical workflow:

1. Check FP3 results before making qualifying predictions. Use the classification as your starting point. The qualifying prediction guide covers what to look for.
2. Use qualifying results as your race prediction baseline. Don't shuffle more than 2-3 positions unless you have a strong reason. The race prediction guide covers what justifies moving someone.
3. On sprint weekends, weight sprint session results more heavily than practice data.
4. Be skeptical of your own adjustments. Every time you move a driver, ask yourself: "Is this based on data, or do I just feel like this driver is due for a good result?"

The bots in this experiment prove that F1 session data has genuine predictive value. The practice tools and session analysis built into Podium Prophets give you access to exactly this data. The question isn't whether data helps. It clearly does. The question is whether you can layer human judgment on top of it without making things worse.

Based on this experiment, the answer is yes. But only if you're disciplined about it.

The full strategy breakdown

For anyone who wants to replicate this experiment or build on it, here's exactly what each bot did:

FP1 Freddie used Friday morning FP1 results for both qualifying and race predictions. Every weekend, same rule. This is the "minimum possible effort" strategy and it shows. FP1 data is noisy and exploratory.

Lazy Lando used FP3 results for qualifying and the starting grid for race predictions. On sprint weekends, he used sprint race results instead of FP3 for qualifying. This is what your friend does when they run out of time and just copy the most recent session.

Quali Quinn used FP2 results for qualifying predictions (sprint qualifying on sprint weekends) and main qualifying results for race predictions. The race prediction is essentially "predict zero overtakes."

Sprint Sage matched Lazy Lando on normal weekends but used sprint qualifying for qualifying predictions and sprint race for race predictions on sprint weekends. The purest test of whether competitive sprint data beats practice data.

Standings Stan used the cumulative championship standings entering each weekend for both predictions. No weekend-specific data at all. Skipped Round 1 (no prior standings).

Last Race Larry used the previous race's finishing order for both predictions. Pure momentum play. Skipped Round 1 (no previous race). Finished near the bottom because F1 form is circuit-dependent, not sequential.

---

I built Podium Prophets because my friends and I were tired of tracking predictions in a spreadsheet. It scores your qualifying and race picks automatically and has built-in session analysis so you can see exactly how FP3 pace translates to race day. Free to use, made by a fan for fans.