<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Daniel Tobi on Medium]]></title>
        <description><![CDATA[Stories by Daniel Tobi on Medium]]></description>
        <link>https://medium.com/@DanielTobi0?source=rss-90831fb9d89e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*NDrM6yOeq2Nn86S58i5oaA.png</url>
            <title>Stories by Daniel Tobi on Medium</title>
            <link>https://medium.com/@DanielTobi0?source=rss-90831fb9d89e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 16:08:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@DanielTobi0/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Automating Football Analysis with Effective Context Engineering]]></title>
            <link>https://medium.com/@DanielTobi0/automating-football-analysis-with-effective-context-engineering-5919385ea221?source=rss-90831fb9d89e------2</link>
            <guid isPermaLink="false">https://medium.com/p/5919385ea221</guid>
            <category><![CDATA[context-engineering]]></category>
            <category><![CDATA[football-betting]]></category>
            <dc:creator><![CDATA[Daniel Tobi]]></dc:creator>
            <pubDate>Sat, 02 May 2026 11:44:11 GMT</pubDate>
            <atom:updated>2026-05-02T11:44:11.567Z</atom:updated>
            <content:encoded><![CDATA[<p>I’ve been working on a project lately, football analysis with effective context engineering.</p><p>Football is considered to be the sport with the highest level of uncertainty. Still, with careful analysis, you will always have an edge that makes you break even and be profitable in the long run.</p><p>And that’s why I’m automating the hardest part.</p><h3>Project Goal</h3><p>automate the most time-consuming part of football analysis: collecting, structuring, and interpreting match context across multiple stages.</p><p>It produces three analysis moments for each fixture:</p><p>1. Pre-match analysis</p><p>2. +20-minute live analysis</p><p>3. +75-minute live analysis</p><h3>Context Engineering for Teams</h3><p>To evaluate a match, I had to come up with these strategies.</p><p>1. Match metadata</p><ul><li>Date</li><li>Home/away venue</li><li>Half-time and full-time scorelines</li><li>Result (win, draw, loss)</li></ul><p>2. Chance quality and shot profile</p><ul><li>Expected goals (xG), including first-half and second-half xG</li><li>xG against and non-penalty xG (npxG)</li><li>Total shots, shots on target, shots in the box</li><li>Big chances and big chances missed</li></ul><p>3. Possession and territory control</p><ul><li>Penalty area touches</li><li>Final-third entries</li><li>Corner kicks</li></ul><p>4. Buildup and progression</p><ul><li>Accurate crosses</li><li>Accurate long balls</li></ul><p>5. Defensive contribution</p><ul><li>Clearances</li><li>Interceptions</li><li>Goals prevented</li></ul><p>6. Team characteristics gotten from whoscored[dot]com</p><h3>Pre-Game Analysis Pipeline</h3><p>The pre-game stage combines multiple sources:</p><p>1. Historical team data from the stats provider</p><p>2. Team characteristics extracted from WhoScored pages</p><p>3. League table grounding for current season position/context</p><p>The output is a pre-match report focused on likely match events and conditions rather than deterministic final-score claims.</p><h3>In-Game (Live) Analysis Pipeline</h3><p>After kickoff, the system schedules and runs two live analysis updates.</p><p>1. +20 minute update</p><p>- Captures early tactical intent and game-state direction</p><p>- Uses live match stats (for example, xG, possession, shots on target and fouls), including the pre-game analysis contexts.</p><p>2. +75 minute update (realistically it’s 50–60mins)</p><p>same as the 20mins, but executed in 75mins after the match begins.</p><h3>Confidence Framework</h3><p>Every key event-oriented prediction is tagged with one of four confidence labels:</p><p>1. Very Likely</p><p>2. Likely</p><p>3. Possible</p><p>4. Unlikely</p><p>These labels are intentionally categorical, not numerical, due to the non-deterministic nature of LLM.</p><p>How to use the service</p><p>It can be accessed from this URL</p><p><a href="https://frontend-football-prediction.onrender.com/">https://frontend-football-prediction.onrender.com</a></p><h3>What didn’t work for me?</h3><p>Fundamental analysis (news, injuries, suspensions)</p><p>I focused a lot on recent developments, injuries, suspensions, and team news. But in reality, for top clubs, these factors often have less impact than expected. There’s almost always a capable replacement, so the absence of even a key player doesn’t shift outcomes as much as I assumed.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5919385ea221" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top-p sampling in simpler terms.]]></title>
            <link>https://medium.com/@DanielTobi0/top-p-sampling-in-simpler-terms-4a209f5e147a?source=rss-90831fb9d89e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a209f5e147a</guid>
            <category><![CDATA[gpt]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Daniel Tobi]]></dc:creator>
            <pubDate>Fri, 03 Jan 2025 07:07:33 GMT</pubDate>
            <atom:updated>2025-01-03T07:08:53.021Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="an image of top-p sampling" src="https://cdn-images-1.medium.com/max/259/1*9r0NS3ezw7gDz_Ocl5ISow.png" /></figure><p>top-p sampling is a sampling method in which you consider only the smallest set of tokens whose cumulative probabilities sum to a value less than or equal to the top-p value.</p><p>This means that the tokens whose cumulative probabilities sum to p or less are considered for sampling.</p><p><em>jargons right? …..a detailed explantion below.</em></p><h3>example</h3><p>let’s use a list containing some probabilies values.</p><blockquote>sample probabilities = [0.15, 0.04, 0.40, 0.06, 0.25, 0.10]</blockquote><blockquote>top-p value(threshold) = 0.90</blockquote><p><strong>steps in top-p sampling</strong></p><ol><li>sort the sample probabilities from highest to the lowest.</li></ol><blockquote>sample probabilities = [0.15, 0.04, 0.40, 0.06, 0.25, 0.10]</blockquote><blockquote>sorted sample probabilities = [0.40, 0.25, 0.15, 0.10, 0.06, 0.04]</blockquote><p>2. calculate the cumulative sum of the sorted sample probabilities</p><blockquote>cumulative sum of sample probabilities = [0.40, 0.65, 0.80, 0.90, 0.96, 1.0]</blockquote><blockquote>let’s call this cumsum_probs</blockquote><p>3. filter the tokens based on the top-p value.</p><p>Remember:</p><blockquote>sorted sample probabilities = [0.40, 0.25, 0.15, 0.10, 0.06, 0.04]</blockquote><blockquote>cumsum_probs = [0.40, 0.65, 0.80, 0.90, 0.96, 1.0]</blockquote><blockquote>top-p value(threshold) = 0.90</blockquote><p>by filtering we are then left with the first 4 tokens, because their cumulative probabilities are within or less than top-p value (0.90).</p><blockquote>tokens within top-p = [0.40, 0.25, 0.15, 0.10] (corresponding to cumulative probabilities [0.40, 0.65, 0.80, 0.90])</blockquote><p>4. after filtering, we then need to renormalize the selected probabilities values to 1.</p><blockquote>sum of selected probabilities = 0.40 + 0.25 + 0.15 + 0.10 = 0.90</blockquote><blockquote>renormalized probabilities = [0.40/0.90, 0.25/0.90, 0.15/0.90, 0.10/0.9]</blockquote><blockquote>final probabilities = [0.44, 0.28, 0.17, 0.11]</blockquote><p>If you have any questions or find an error here, do not hesitate to let me know.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a209f5e147a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Climbing the Zindi leaderboard]]></title>
            <link>https://medium.com/@DanielTobi0/climing-the-zindi-leaderboard-3390c5086add?source=rss-90831fb9d89e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3390c5086add</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Daniel Tobi]]></dc:creator>
            <pubDate>Wed, 10 Apr 2024 23:33:06 GMT</pubDate>
            <atom:updated>2024-04-11T02:00:40.976Z</atom:updated>
            <content:encoded><![CDATA[<p>I looked at the winning solutions from Zindi’s Tabular dataset competitions and analyse the top method used by the participants.</p><figure><img alt="A logo of Zindi" src="https://cdn-images-1.medium.com/max/318/1*7a3lgUPgkhtPMeXe58oRjA.png" /></figure><p>I will be sharing them in this article.</p><p>1. Data Cleaning:</p><p>This serves as the backbone of their overall model performance; they spend a great ton of time cleaning the data. Techniques used for data cleaning were.</p><p>i) Converting features to the appropriate data type.</p><p>ii) Dropping features with only a single unique value, as they do not contribute to the model learning.</p><p>iii) Removal of outliers. While outliers can sometimes cause overfitting or bring bias, some winning solutions did not bother to remove them, or maybe they never really had a negative impact on the model performance overall.</p><p>iv) Being creative when filling for missing values. Even though some ensemble models like CatBoost can deal with missing values, some brilliant solution simply mapped missing values to ‘UNK’ a placeholder for missing values.</p><p>v) Training the model on part of the data without missing values to predict and fill in those that contains missing values, a strategy that proved to be effective for some participants.</p><p>2. Feature Engineering</p><p>While data cleaning creates the backbone to an effective model performance, feature engineering will make a big difference across competitor’s solutions. It is very much possible for competitors to have a similar approach in cleaning the data, but the feature engineering will almost never be the same. The key aspect of feature engineering I found include.</p><p>i) Feature mapping, which involves extracting additional information from features. For example, categorising numerical ranges into discrete values.</p><p>ii) Feature manipulation, in this approach combination of multiple features or single feature, goes through some mathematical manipulation. Although there is no popular rule to this, you just have to experiment on which feature manipulation will increase the model performance, while some might even decrease the performance or have no effect at all.</p><p>iii) Datatime features, when given features with datetime objects, they have a lot of features that can be squeezed out of them, such as day of week, months, year and so on.</p><p>iv) Cyclic method, when we represented with features like the days of the week in numbers (0–7), we found out that when modelling the model associate the closeness of 0 (Monday) to 2 (Wednesday), but in reality, Monday (0) is closer to Sunday (7) than 3 (Wednesday), so when we use the cyclic method, this brings the values that should be closer to each other in reality closer, and also for the model to better understand their relationships.</p><p>v) Group-by function to aggregate data based on specific features and extract additional insights.</p><p>vi) Shifting, Lag and Rolling, method used for time series analysis and data manipulation. <a href="https://www.linkedin.com/in/olufemi-victor-tolulope/">Professor</a> used this in conjunction with the group-by function in a fantastic way, and it really worked out well for him. A top notched expert.</p><p>3. Modelling</p><p>From this point, there is usually a few unique things to do here, but still a great deal.</p><p>i) Cross Validation, to train the model on a subset of the data and evaluate on the remaining part of the data. This is used to avoid overfitting, leading to a confident model on unseen data.</p><p>ii) Commonly used modelling algorithm were XGBoost and CatBoost.</p><p>iii) Utilizing tool like Optuna for hyperparameter tuning.</p><p>iv) Ensemble methods to combine outputs from different models to enhance the overall predictive model performance.</p><p>Last but not the least LUCK!!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3390c5086add" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>