We gave our Symnasium agent the 2024 MLB season data with a simple task:
Forecast the next 30 games of Dodgers performance starting from May 11, 2024. No per-team training. No manual tuning. Just: here’s the data, predict what happens next.
The Data We Gave It
What We Asked
Game-to-game runs are almost random. But a team’s form — its 10-game rolling run average — moves with real momentum: hot streaks and cold stretches you can see and act on. The question is whether a model can forecast that trend accurately over a full month.
Symnasium forecast the Dodgers’ 10-game rolling run trend over the next 30 games to a 10.45% WAPE (weighted error). It tracked the team cooling off mid-window and heating up down the stretch — not just guessing the season average.
How does that compare?
Why this matters: If you're trying to read a hot streak — is it real, or a mirage? Is it the power surge, the walks, or the base hits that's actually driving it? — the data gives you a clear answer. A rising hit rate is the strongest leading signal that a scoring trend will hold; home runs and walks move the needle far less. Get on base with hits, and the scoring trend follows.
We handed the identical task to a general-purpose Claude agent running Opus 4.8 on Ultracode that built its own time-series model. Its best model landed at 21.32% WAPE, roughly 2× less accurate. The chart tells the story: Symnasium follows the trend; the from-scratch baseline flatlines at the season average and misses every swing.
Figure: Symnasium's rolling run trend forecast (10.45% WAPE) versus an Opus build model (21.32% WAPE) on Dodgers' 2024 season data.
What drives the trend?
Recent hits are the strongest signal — rolling hits correlate most tightly with the rolling run trend (0.73), ahead of walks (0.58) and home runs (0.33). Get on base with hits, and the runs follow.
The value of the right covariates
Forecasting the trend from the run history alone gets you to ~27% WAPE. Adding the recent-form covariates cuts that to 10.45% — a ~60% reduction in error. That’s a concrete answer to “which data feeds are worth paying for."
The Tagline:
Team form (10-game rolling runs): 10.45% forecast error (WAPE) — about 2× more accurate than what was returned by Claude Opus 4.8.
It took seconds, with no per-team training and no manual tuning.
Recent-form covariates cut error ~60% — the platform quantifies exactly which inputs matter.
~60% error reduction with the right recent-form covariates
Here’s the most important part: Symnasium didn’t just give us a forecast — it told us how much to trust it.
The agent ran an automatic calibration audit on the form forecast and found the uncertainty bands were overconfident: the model’s “80% confident” range actually caught the true value only about 57% of the time — the bands were too narrow.
What does this mean?
The model was saying "I'm 80% confident the true value will fall in this range"—but in reality, the bands were too narrow. The model caught outcomes less often than the nominal confidence level suggested.
Sharper than guesswork: Symnasium forecast the Dodgers' next 30 games of scoring form to 10.45% WAPE — about 2× more accurate than Claude Opus 4.8 forecasting on its own (21.32%). On a trend that shapes trade, bullpen, and lineup calls, that's the difference between reacting to noise and reacting to signal.
Driver discovery: Recent hits are the #1 signal of scoring form — correlating more than twice as strongly with the trend as home runs, and clearly ahead of walks. Prioritize contact and on-base skills over power-only bats when you're trying to sustain — or predict — a hot streak.
Set betting lines with honest uncertainty. The agent found uncertainty bands were overconfident: the model’s "80% confident" range actually caught the true value only about 57% of the time. Get the full range of outcomes (blowouts to shutouts), not just the average.
Know what your data is actually worth
Data vendors promise better forecasts. But by how much?
What we measured: The right data cut errors by 50%
What this means: Only pay for data that actually improves your forecasts (proven, not promised)
Catch bad data before it costs money
Forecast attendance and plan accordingly
Using last year's average for staffing and inventory creates waste.
With Symnasium: Forecast attendance using weather, opponent, and team momentum — the same trend-tracking approach that forecast Dodgers scoring form to 10.45% WAPE, applied to demand instead of runs.
What this saves: ~15-25% less food waste, ~10-15% lower overstaffing costs
Forecast hundreds of teams and metrics weekly
Traditional approach: Build a separate model for each (slow, expensive, needs a data team).
With Symnasium: One model handles all 900+ forecasts in seconds
What this saves: ~90% less time, ~80% lower costs.