MonteEntropy - Time Impact Explorer

Introduction

Chess players use engines to help evaluate positions. While this information is very useful, it is limited to evaluation in pawns (e.g., +3.2, -1.17) and ignores other important aspects of the game such as time pressure, player strength, and the stage of the game. and therefore this information is of limited help in accessing practical advantage or disadvantage a player had during a game.

Among these, time pressure plays a particularly important role, especially at shorter time controls, yet it is not captured by traditional engine evaluations, and therefore this information is of limited help in accessing practical advantage or disadvantage a player had during a game.
Time Impact Explorer is a tool that attempts to address this gap. It uses machine learning to provide, in addition to engine evaluation, an estimate of the outcome of the game (win / draw / loss), taking into account factors such as:

• Engine evaluation
• Remaining time on the clock
• Player ratings
• Game context

How it works / How to use it

Just pass any lichess game id or game url, and click "Compute Expected Outcome Probabilites", wait for the model to compute and you will see something like this:

The most interesting chart is the "Outcome Probabilities" this is where our model estimates the outcome of the game. Notice that it is different from what the evaluation bar says, according to the evaluation bar black is winning (-1.77). However, since white is much better on the clock 1:27 to 0:05, the estimated probability of white winning is 59.8% / draw 16.1% / black winning 24.1%.

Intuitively, the given by: "In a similar positions between players of similar strengths, and similar time situations this is what happened".
One can double-click on the slider to see forces that are pushing toward each estimation.

The biggest force pushing towards white win (and away form black win) is Time, biggest force pushing away from white win and towards black win is Position.

Methodology

The core idea behind Time Impact Explorer is simple: instead of evaluating positions purely from an engine perspective, we attempt to model what actually tends to happen in real games.

To do this, we trained a machine learning model on approximately 420,000 positions extracted from real Lichess games. For each position, the model attempts to predict the eventual outcome of the game (win / draw / loss).
The model uses features such as:

• Engine evaluation
• Remaining time for both players
• Player ratings
• Material and positional information

Once trained, the model is applied to every position in a game. For each move, we compute the relevant features and use the model to estimate the probabilities of White winning, drawing, or Black winning. These estimates are then displayed in the Outcome Probabilities chart.

Limitations / Future Work

1. Data Coverage / Statistical Variability

The model can only learn from the positions and situations present in the training data. Certain time controls, rating ranges, and position types may be underrepresented. Like most machine learning systems, the model may also occasionally produce unintuitive estimates or react unexpectedly to small changes in the position or game situation.

Future work includes expanding the dataset, incorporating more targeted training data, and improving the overall stability and consistency of predictions.

2. Engine Evaluation vs Practical Human Evaluation
Stockfish evaluation does not always correspond perfectly to practical human advantage. Two positions with similar engine evaluations may be dramatically different from a human perspective: one position may be straightforward to convert, while another may require finding only moves under severe time pressure.

Future work includes experimenting with more human-oriented evaluation signals, such as Maia evaluations, as well as incorporating additional practical difficulty metrics.

MonteEntropy - Time Impact Explorer