tactica. already produces Match Intelligence and recommendations. But recent auditing has pushed us towards a more fundamental question: what should a football model actually understand before it has earned the right to form an opinion about a match?
tactica. already makes football predictions.
That sounds like the important part.
Increasingly, we don't think it is.
The more we've audited the system—its evidence, explanations, recommendations and the assumptions sitting underneath them—the more we've found ourselves asking a different question.
What should a football model actually understand before it makes a prediction?
Not how many statistics it can collect.
Not how complicated its mathematics can become.
Not how confidently it can describe the output afterwards.
What does it need to understand about the football match that is actually going to be played?
That question is beginning to shape the next stage of tactica.
And it starts well before the recommendation.
The team in the database isn't necessarily the team playing tonight
Imagine a football team that produced excellent attacking numbers last season.
Goals. Chances. Expected goals. Shots. Territory.
Those numbers are real.
They happened.
They belong in the historical evidence.
Now imagine that several of the players responsible for creating that attacking output leave during the summer.
The historical numbers haven't suddenly become false.
But something important has changed:
How representative are those numbers of the team that will play tonight?
A football club is a continuous institution.
A football team isn't.
Players leave. Players arrive. Managers change. Roles change. Injuries alter combinations. Formations move. A player who generated a large amount of attacking output might be replaced by someone with a completely different profile.
That creates a problem for any model heavily dependent on historical team statistics.
Last season's team remains useful evidence.
It shouldn't simply be treated as today's team.
The standard we're now working towards is closer to this:
Historical team evidence establishes the prior.
Current player and squad evidence tells us how representative that prior still is.
The expected lineup refines our view of the team likely to play.
The confirmed lineup eventually replaces assumption with fact.
This is important context for what follows: much of this article describes the intelligence architecture we're working towards, not capabilities we're claiming tactica. already has in production. Today tactica. has a functioning football model, Match Intelligence, recommendations, Case Files and Performance foundations. The next stage is about earning substantially deeper football understanding before those capabilities reach the live product.
Before the team sheet, we have assumptions
Lineups are a good example of the distinction we're trying to make.
Before an official team sheet exists, tactica. should eventually be capable of forming its own expected XI from factual evidence.
Who belongs to the current squad?
Who is available?
Who has been starting?
How many minutes have they played?
What positions have they occupied?
What formations has the team been using?
Who normally replaces whom?
That produces an assumption.
Hopefully a good one.
But still an assumption.
Then the official lineup arrives.
Now we have fact.
The interesting part isn't simply scoring tactica. on whether it predicted nine, ten or eleven starters correctly.
The better question is:
Did the difference between our expected team and the actual team materially change our understanding of the match?
Sometimes it should.
Imagine tactica. expected a team's primary creator to start and built its attacking expectation partly around that assumption.
The team sheet arrives.
The player is on the bench.
Their replacement has materially weaker validated creative evidence.
Now there may be a legitimate reason for the expected attacking output to change.
But sometimes a predicted XI will be wrong and almost nothing meaningful should happen.
One full-back replaces another with a broadly comparable profile.
The expected lineup changed.
The football expectation didn't materially change.
That's also an intelligent conclusion.
We don't want the model to manufacture movement just because new information arrived.
We want it to understand whether that information matters.
A match isn't two spreadsheets
This leads to another problem.
A lot of football analysis—automated or otherwise—essentially does this:
Team A statistics
versus
Team B statistics
There is value in that.
But it isn't quite the same as modelling the match.
Football matches are interactions.
Players encounter particular players.
Units encounter particular units.
One team's preferred way of attacking meets another team's preferred way of defending.
Game states alter behaviour.
A side protecting a lead behaves differently from one chasing a goal.
So one of the questions we're now investigating is whether tactica. can move from simply understanding two teams towards understanding what happens specifically when those teams meet.
Consider bookings.
A defender might have a relatively high historical card rate.
Useful information.
An opposing winger might attempt lots of take-ons and regularly draw fouls.
Also useful.
But the more interesting football question is:
Is this particular match likely to repeatedly put that defender into the situations that cause them to foul?
Perhaps the defender's role leaves them isolated against that winger.
Perhaps the opponent's attacking pattern repeatedly directs play towards that side.
Perhaps the expected game state makes those encounters more likely.
That's a richer football hypothesis than:
Defender averages X cards per 90.
But richer language does not automatically mean richer intelligence.
Football-sounding reasoning still has to earn its place
This is where we need to be careful.
It is remarkably easy to make a football model sound intelligent.
Take an older centre-back and a quick winger.
You could write:
The defender's lack of pace could leave him vulnerable against the winger.
Sounds plausible.
What evidence actually established that the defender lacks pace?
Age?
That's not enough.
Likewise:
Height isn't aerial ability.
Position isn't tactical role.
A winger isn't automatically a strong dribbler.
If tactica. can't establish something factually, it doesn't get to manufacture the missing evidence because the resulting sentence sounds like good football analysis.
That creates a much harder standard for the Matchup Intelligence we're investigating.
A football hypothesis needs an authoritative factual basis.
Then it needs to become something measurable.
Then we need historical evidence that the mechanism actually matters.
Then it needs to survive data it wasn't developed against.
Only after that should we consider allowing it to influence production Match Intelligence.
And even then, we need to preserve what it contributed so we can judge it afterwards.
Some ideas will fail that process.
We expect them to.
A mechanism we thought mattered might add no useful signal.
Another might work historically and disappear out of sample.
Something that seems obvious to anyone who watches football might turn out to be too difficult to represent reliably with the evidence available to us.
Rejecting those mechanisms isn't failed development.
It's part of building the model.
The recommendation comes last
All of this changes how we think about recommendations.
The temptation with a betting-oriented football product is to begin with the market:
What's the bet?
We're increasingly trying to work in the opposite direction.
First establish reality.
Then establish what evidence we actually possess.
Then understand the teams and players.
Then form assumptions about who is likely to play.
Then understand how those teams might interact.
Then form the expected match state.
Then simulate uncertainty.
Then decide whether anything has earned qualification.
The recommendation is the end of the reasoning chain.
Not the beginning.
That distinction also means tactica. needs to be comfortable reaching an answer that isn't commercially exciting:
NO RECOMMENDATION.
Sometimes understanding the match should produce no bet
Not every fixture needs a recommendation.
Not every market needs an opinion worth acting on.
The evidence might be weak.
Different evidence might contradict itself.
Lineup uncertainty might be too important.
The model's probability distribution might be too broad.
The available price might not justify execution.
A market might simply fail the evidence standard required for tactica. to publish it.
None of those outcomes means tactica. failed to analyse the match.
In fact, recognising them may be evidence that the analysis worked.
There is another distinction here that matters:
A football opinion isn't automatically a betting recommendation.
tactica. might conclude that Arsenal are the most likely team to win a particular match.
That doesn't mean any available Arsenal bet has earned qualification.
The model can have an opinion about the football while having no recommendation about the market.
If the reasoning chain ends there, the product should be able to explain why.
We don't want to build this directly into production
There is an obvious danger in everything I've described.
It sounds good.
Player-aware team intelligence.
Predicted lineups.
Confirmed-lineup recalculation.
Matchup Intelligence.
Deeper reasoning.
It would be very easy to turn those concepts into a list of features and start adding them to the current product.
We're deliberately not doing that.
The next generation of Match Intelligence is intended to be developed away from the existing production model.
Current tactica. continues doing its job.
New intelligence mechanisms get developed and tested separately.
If they survive historical testing and out-of-sample validation, they can eventually operate in shadow alongside production on real upcoming football matches.
The shadow system can make its pre-match assumptions.
Those assumptions can be frozen.
Official lineups can challenge them.
Matches can happen.
Then we can evaluate what the new intelligence actually understood.
Only evidence should earn promotion into production.
There is no useful launch date for that process.
If an idea needs another month of validation, it needs another month.
If something fails, it fails.
A calendar doesn't get to make an intelligence mechanism trustworthy.
The interface comes afterwards
This has also changed how we're thinking about the next version of tactica.'s Case File.
It would be easy to design a beautiful new interface now.
Create sections for lineups.
Create a Matchup Intelligence panel.
Build graphs showing expected-goals movement.
Add animations when a team sheet changes.
Then ask the model to fill everything we've designed.
We're increasingly convinced that's backwards.
First determine what tactica. can actually understand.
Build it.
Test it.
Reject what doesn't survive.
See which intelligence genuinely changes our understanding of football matches.
Then design the Case File around those truths.
One match might be dominated by squad turnover.
Another by a meaningful player matchup.
Another by contradictory evidence.
Another by a confirmed lineup that materially changes the expected match state.
And another might ultimately deserve a very simple conclusion:
No Recommendation.
The intelligence should earn the interface—not the other way around.
What we're actually trying to build
None of this means historical team statistics stop mattering.
They matter enormously.
It doesn't mean every player needs a complicated individual model.
It doesn't mean every tactical interaction needs to become a feature.
And it certainly doesn't mean adding more data automatically makes tactica. more intelligent.
The direction is almost the opposite.
We want tactica. to become more disciplined about understanding what each piece of evidence is entitled to tell us.
Historical evidence establishes context.
Current evidence challenges whether that context still applies.
Assumptions remain assumptions until facts replace them.
Football hypotheses remain hypotheses until evidence validates them.
New intelligence remains outside production until real testing earns promotion.
And after all of that, sometimes the correct decision should still be:
Do nothing.
That's a substantially harder system to build than one which simply produces more predictions.
We think it's also a more interesting one.
The goal isn't to make tactica. sound more intelligent.
It's to make the reasoning underneath it more deserving of the word.