Me and DeepSeek analyzed 1600 Codeforces contests and 8000+ problems — some interesting patterns

please check this result : https://ayushgirigoswami.github.io/codeforces_analysis_report/

A few weeks ago I got curious about something most of us probably notice intuitively but rarely measure properly:

• Is Codeforces getting harder over time?
• Which topics are becoming more common?
• Are some problems overrated or underrated?
• How different are Educational rounds from regular rounds?

So I wrote a Python script that pulls contest/problem data from the CF API, analyzes ratings + tags, and generates interactive visualizations.

Report

https://ayushgirigoswami.github.io/codeforces_analysis_report/

Source code

https://github.com/Ayushgirigoswami

Dataset

The analysis includes:

• 1600 contests
• 8354 rated problems
• Div.1 / Div.2 / Div.3 / Div.4 / Div.1+2
• 2011 → present

Average problem rating across the dataset: 1793

Some interesting results

1) Div.1+Div.2 rounds have the widest difficulty spread

Div.1+2 rounds have by far the largest standard deviation.

Makes sense in hindsight: these rounds combine easy entry problems with very high-end G/H problems, so the spread becomes huge.

Also interesting: even Div.1 A problems average around 1537, which is already harder than many Div.2 mid-problems.

2) Biggest difficulty jumps are usually B→C and C→D

Average Div.2 ratings by position:

• A ≈ 903
• B ≈ 1203
• C ≈ 1552
• D ≈ 1932
• E ≈ 2300
• F ≈ 2614

The largest jumps are:

• B → C : +349
• C → D : +380

This matches what many contestants experience during contests: B is often straightforward, while C/D is where actual problem solving starts becoming important.

For Div.1+2 rounds, the jump near the end becomes even more extreme:

• F ≈ 2697
• G ≈ 3102
• H ≈ 3160

3) Topic trends over time

Increasing frequency

• greedy
• math
• constructive algorithms
• data structures
• binary search
• dp
• trees
• bitmasks
• interactive

Decreasing frequency

• implementation
• geometry

The increase in interactive problems during the last few years was especially noticeable.

Geometry also appears much less frequently than older rounds.

4) Most common tags

Some observations:

• DP is disproportionately common in Div.1.
• Implementation dominates Div.2/3 but drops heavily in Div.1.
• Graph-related problems appear much more frequently than I expected.

5) Educational rounds vs regular rounds

This part surprised me.

Overall average difficulty is almost identical.

But position-by-position: Educational rounds are consistently easier from A→E, while F problems are actually harder on average.

6) Problems whose ratings seem unusual

Examples of problems that appear easier/harder than their ratings suggest based on solve counts.

Easier than expected

• 1264F — Beautiful Fibonacci Problem (rated 3500, but solved by 1000+ users)

Harder than expected

• 2190F — Xor Product
• 2066F — Curse
• 1967F — Next and Prev
• 949F — Astronomy

These had surprisingly low solve counts relative to their ratings.

7) Contest “symmetry”

I also tried measuring how balanced contest difficulty curves are.

Average symmetry score: 0.454 / 1

Interpretation:

• 0.7 → balanced progression

• < 0.4 → heavily front-loaded

Most CF contests lean slightly front-loaded: easy opening problems followed by a sharp wall.

Running the script

Requirements:

pip install requests pandas numpy plotly matplotlib tqdm

Run:

python deep.py

The script:

1. Fetches contests/problems from the CF API
2. Performs statistical analysis
3. Generates interactive Plotly graphs
4. Detects trends/anomalies

Fetching everything takes around 15–20 minutes because of API rate limiting.

One thing I still want to analyze is solve timing during contests (for example: when most users solve C/D problems), but that would require a much larger amount of contest.status API calls.

If anyone has ideas for additional analyses, suggestions are welcome :)