DOTA2PROTIPS

How it works/Predictions

How I Make My Dota 2 Predictions

What this is

Dota2ProTips is my personal prediction log. I — Greg Spencer — watch Dota, and before each series starts I put a pick on the line: who I think is going to win, and why. That's the whole idea.

Every prediction is public, every result is tracked, and my accuracy is visible to anyone who visits. I can't quietly delete a bad call. That's the point.

The model and me

Behind the picks sits an internal model — ELO team ratings combined with recent results and momentum signals — that gives me a baseline win probability for each match. It's a useful discipline tool, a sanity check against my own biases. But it's a starting point, not the answer.

The model is great at the macro picture. It can't account for specific matchup dynamics, current form, or the micro-realities of a given match. That part is on me. How much is me and how much is the model depends on the tier, below. For a longer write-up of how I approach a match, read this.

The one rule: before the draft

Every pick is locked before hero selection. By the time the draft starts, it's already written down. That matters, because the draft in Dota reveals a huge amount of information — predicting after it is far easier and far less honest. Once a pick is published it never changes. If I was wrong, it stays wrong on the site.

Tier 1 — the analyzed picks

The big events I watch in full and write every call myself. The model advises, I decide. What goes into it:

  • Recent form — how each team has played in the last 2–4 weeks. Momentum matters more than historical prestige.
  • Playstyle matchup — how each team wants to win, and whose plan is more likely to work against the other. Fast vs. slow, skirmish vs. farm — the interaction matters more than raw rating.
  • Draft tendencies — narrow hero pools, patch comfort, drafting habits. I never flip a pick after seeing the draft, but long-term drafting trends are factored in before the match.
  • Head to head — how these two teams have matched up before — only when the roster and meta are comparable.
  • Roster changes — stand-ins, bootcamps or roster shifts that could affect performance.
  • Tournament stage — groups play differently than playoffs. Teams sometimes sandbag in groups to hide strategies.
  • ELO rating — my rating system gives a structured view of where each team stands relative to the other.

After the match I come back and write an aftermath — a short note on what actually happened, what I got right, and where I was wrong. Predictions without follow-up accountability are worthless.

Tier 2 and qualifiers — the quick picks

There are hundreds of these matches, and hand-writing an analysis for every one isn't realistic — so this is where the model earns its keep. My internal system does the first pass: it reads the database — teams, players, recent results, head to head, ratings — and weighs around 62 parameters across form, matchup, roster, ratings, patch and stage into a probability for each match.

I read that output, sanity-check it, and lock a pick — a winner, or a draw for a Bo2. When I know something the numbers don't, I override it. No essay on every match, and no pretending there is one. Just an honest, tracked call, made the same way every time.

The probability bar and upset calls

Every match shows a bar with the ELO model's chances (how they're calculated). For a Bo2 it carries three outcomes — Team A win, draw, Team B win — so a close match shows a fat draw block in the middle.

When I back a team whose per-game chance is at least 4 points below its opponent's, that's an upset call, and it's marked on the card. When one lands, the card turns gold. So far there have been 164; 97 landed.

How accuracy is scored

  • A pick is correct if the chosen team wins the series.
  • A 1–1 Bo2 draw counts as correct only if I picked the draw.
  • Accuracy = correct / (correct + wrong). Pending matches don't count.
  • Tier 1 and quick picks are tracked separately — an analyzed Tier 1 call and a quick model pick aren't the same product, and blending them into one number would be dishonest.

Right now: 934 picks, 68.9% over 916 settled. Tier 1 is at 67.8% (603), quick picks at 70.9% (313). On the same series, always backing the ELO favourite would have scored 65.9% against my 69.5%.

I aim to beat 60% on any given event. Sounds modest, but in a game where the best teams in the world lose on any given day, staying above a coin flip is genuinely hard. Everything is stored openly on the track record.

Series MVP

After some series I name a Series MVP. It's not a single stat — it's a weighted scorecard. The OpenDota numbers give the raw picture, my model adds a sanity check, and the final call is mine. If there's no clear MVP, it stays empty — I don't force one just to fill the slot. What carries the most weight, roughly in order:

  • Impact on wins — who was most responsible for the decisive moments — game-winning fights, key initiations, clutch saves, or vision that let cores free-farm.
  • Consistency across the series — solid across most games beats one godlike game followed by a disappearing act.
  • Role-adjusted performance — cores on hero damage, net worth and clean KDA; offlaners on initiation and space; supports on disables, wards and hero-specific plays.
  • Objectives and map control — towers, barracks, Roshan, runes and vision.
  • Hidden impact metrics — kill participation, damage and healing share, crowd-control duration, GPM/XPM relative to role.
  • The eye test — sometimes one player simply feels like the difference-maker.

Not betting advice

Picks are my opinion and analysis, made for entertainment. The details are in the terms of use.