Some simpler things I think would be interesting:

• Who has done the most damage/healing/kills/ etc
  
• Who has the highest/lowest winrate (must have played 100 games or so)
  
• Who has the most time/games played from logs.tf
  
• Who has sent the most text chats from logs.tf
  
• Which classes have the most "net damage" per heals received
  
• Which classes have the highest/lowest damage done per kill
  
• Distribution of match times throughout the day

Slightly more complicated:

• Heatmap of net damage
  
• Heatmap of uber drops (both med and attacker)
  
• Compare stats from "official match days" to other days (esp match length)
  
• "diversity of players" (how many people just play 6s, hl, bball, etc.? how many people play ultiduo but not bball? do people play each class evenly, or do people play just one or two? etc)
  
• You could compare the accuracy per distance to target from cheaters to non-cheaters

I got a lot more ideas, Ill prob msg you via discord to avoid spamming this thread

Probably you don't have enough data to do that for most games but for players that have played loads of times against/with each others (frequent puggers or teams with core of players that have stayed for a long time together) you can try to predict the outcome of a game.
I'm thinking something more advanced than just a simple ELO system where team with higher ELO is more likely to win but rather taking into account classes, players' history with and against each others, maps, etc

But again, you probably don't have nearly enough data for this to be somewhat accurate, except for maybe the top 30 pugchamp (or tf2center ?) players.

I was thinking of doing this with invite players. Seeing how different rosters are predicted (at least statistically) to succeed or fail. But thank you for the suggestions! An ELO system is not something I had thought of.

If you want to do that, you would need to have separate ELO scores for each gamemode/classe IMO because otherwise the data would be so heterogeneous that the ELO would lose all meaning.
For tf2pickup.fr, I have computed an ELO per class per player and since the games are really homogeneous (only 6v6), the resulting ELO scores make sense (for instance, JuJu is by a significant margin bottom ELO as soldier and it completely matches with how impossible it is to win a game when you're matched with him).

But yeah, a general per class per gamemode ELO would also be nice !

Who has the highest net damage (dpm-dtm) in all of tf2

who has played the most matches on each map, my bet for proccess is on NOPE ⅓.

make some thing like this

https://upload.wikimedia.org/wikipedia/commons/d/d2/Internet_map_1024.jpg

of players.

So you could see clusters of players that play a lot together.

highest average med picks/drops by player, then by class

slur heatmap

Time frame players show up in logs (Time in days or years from first recorded log to most recent).

Accuracy (slot1 weapon) graph by player and class (probably for invite only), where X - accuracy percentage (from lowest to highest for every player), Y - number of matches with said accuracy. Too bad most logs from EU servers don't have accuracy stats.

Something like this

how many gigs is the folder of every tf2 log

make a NEET detector which shows which comp players play the most tf2 (i wanna see the sinkarmas of the world with 160hrs/2weeks)

It's about 80 gigabytes right now. I still got about ~390k more logs to download, but that shouldn't increase it by more than a gigabyte at most

you should use this https://github.com/ldesgoui/clone_logs