<?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[The Quantastic Journal - Medium]]></title>
        <description><![CDATA[Follow the Quantastic Journal, where science, technology, and mathematics meet humanity. We publish carefully curated stories that bring rigorous knowledge, new perspectives, and open debate to curious readers who value depth and critical thought. - Medium]]></description>
        <link>https://thequantasticjournal.com?source=rss----cf522ff06849---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>The Quantastic Journal - Medium</title>
            <link>https://thequantasticjournal.com?source=rss----cf522ff06849---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 02:25:02 GMT</lastBuildDate>
        <atom:link href="https://thequantasticjournal.com/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[Scientists Review Each Other’s Work for Free. Publishers Make Billions From It.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/scientists-review-each-others-work-for-free-publishers-make-billions-from-it-e048f181bc1a?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/2600/1*GQ9nbwQUfPOhk4jSNSuXeg.jpeg" width="3840"></a></p><p class="medium-feed-snippet">The best business in science runs on unpaid experts, a 38% profit margin, and lately, secret messages to chatbots</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/scientists-review-each-others-work-for-free-publishers-make-billions-from-it-e048f181bc1a?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/scientists-review-each-others-work-for-free-publishers-make-billions-from-it-e048f181bc1a?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/e048f181bc1a</guid>
            <category><![CDATA[money]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[history]]></category>
            <dc:creator><![CDATA[Ricky Lanusse]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 08:12:34 GMT</pubDate>
            <atom:updated>2026-11-14T08:12:33.083Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Black Holes Without the Full Mathematics of General Relativity: Light Moving Straight In and Out…]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/black-holes-without-the-full-mathematics-of-general-relativity-light-moving-straight-in-and-out-e444fc4418cd?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/2600/0*WiKnBMRYCx3moi1Q" width="5700"></a></p><p class="medium-feed-snippet">About this series: Most engineers never learn the mathematics and physics normally needed to approach studying black holes, and so treat&#x2026;</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/black-holes-without-the-full-mathematics-of-general-relativity-light-moving-straight-in-and-out-e444fc4418cd?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/black-holes-without-the-full-mathematics-of-general-relativity-light-moving-straight-in-and-out-e444fc4418cd?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/e444fc4418cd</guid>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[astronomy]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[cosmology]]></category>
            <dc:creator><![CDATA[Maxwell's Demon]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 14:52:44 GMT</pubDate>
            <atom:updated>2026-11-13T14:52:43.527Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Why Do Galaxies Rotate?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/why-do-galaxies-rotate-513df25194f2?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/1920/1*pg0k_B73iQGvdOiZz2JEjQ.avif" width="1920"></a></p><p class="medium-feed-snippet">Galaxies act a little bit like figure skaters twirling on the ice &#x2014; but at an astronomical scale.</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/why-do-galaxies-rotate-513df25194f2?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/why-do-galaxies-rotate-513df25194f2?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/513df25194f2</guid>
            <category><![CDATA[astronomy]]></category>
            <category><![CDATA[universe]]></category>
            <category><![CDATA[spin]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[astrophysics]]></category>
            <dc:creator><![CDATA[The Conversation U.S.]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 14:47:47 GMT</pubDate>
            <atom:updated>2026-11-13T14:47:45.930Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Harvest Looks Fine. What About The Pollinators?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/the-harvest-looks-fine-what-about-the-pollinators-6b21ca41a7d0?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/2400/1*59_aGozxAhCmMlnYM4a2Yg.png" width="2400"></a></p><p class="medium-feed-snippet">Two major studies published in 2026 find little evidence of a global pollination crisis in agricultural production data. But crop yields&#x2026;</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/the-harvest-looks-fine-what-about-the-pollinators-6b21ca41a7d0?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/the-harvest-looks-fine-what-about-the-pollinators-6b21ca41a7d0?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/6b21ca41a7d0</guid>
            <category><![CDATA[agriculture]]></category>
            <category><![CDATA[environment]]></category>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[biodiversity]]></category>
            <category><![CDATA[science]]></category>
            <dc:creator><![CDATA[Niladri Giri | 10+ Years Agriculture Field Work]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 07:43:56 GMT</pubDate>
            <atom:updated>2026-11-12T07:43:54.844Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Physics of Project Hail Mary: How Far Can Andy Weir Bend Reality?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/the-physics-of-project-hail-mary-how-far-can-andy-weir-bend-reality-9737db147ba2?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/1672/1*jAoH5S4LquERIQwEnYRZ6A.png" width="1672"></a></p><p class="medium-feed-snippet">Photon propulsion, living fuel, neutrinos, and the limits of interstellar travel.</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/the-physics-of-project-hail-mary-how-far-can-andy-weir-bend-reality-9737db147ba2?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/the-physics-of-project-hail-mary-how-far-can-andy-weir-bend-reality-9737db147ba2?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/9737db147ba2</guid>
            <category><![CDATA[biology]]></category>
            <category><![CDATA[scifi]]></category>
            <category><![CDATA[project-hail-mary]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[astrophysics]]></category>
            <dc:creator><![CDATA[Marco Capriz]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 06:48:13 GMT</pubDate>
            <atom:updated>2026-11-12T06:48:11.696Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Geometry of Quantum States]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/the-geometry-of-quantum-states-4844ecc5e300?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/2600/0*5Y5c9a0nVjGVT_Z0" width="2600"></a></p><p class="medium-feed-snippet">Phase, distance, and motion in a space of possibilities</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/the-geometry-of-quantum-states-4844ecc5e300?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/the-geometry-of-quantum-states-4844ecc5e300?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/4844ecc5e300</guid>
            <category><![CDATA[quantum-physics]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[geometry]]></category>
            <dc:creator><![CDATA[Sam Vaseghi]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 19:39:07 GMT</pubDate>
            <atom:updated>2026-11-11T19:39:05.601Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Knight’s Tour: A Mathematical Journey Across 64 Squares]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/the-knights-tour-a-mathematical-journey-across-64-squares-79a18631fd7a?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/1681/1*9yNfXWrxfdNYxArwyGKjkg.png" width="1681"></a></p><p class="medium-feed-snippet">How a knight&#x2019;s journey across a chessboard became a fascinating mathematical problem, connecting a centuries-old puzzle with graph theory&#x2026;</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/the-knights-tour-a-mathematical-journey-across-64-squares-79a18631fd7a?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/the-knights-tour-a-mathematical-journey-across-64-squares-79a18631fd7a?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/79a18631fd7a</guid>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[graph-theory]]></category>
            <category><![CDATA[problem-solving]]></category>
            <dc:creator><![CDATA[Syed Razin Mahmood]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 19:26:25 GMT</pubDate>
            <atom:updated>2026-11-11T19:26:24.141Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Forecasting 48 Hours Ahead: How Much Complexity Was Worth It?]]></title>
            <link>https://thequantasticjournal.com/forecasting-48-hours-ahead-how-much-complexity-was-worth-it-cf4308edce3b?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/cf4308edce3b</guid>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[time-series-forecasting]]></category>
            <category><![CDATA[arima-model]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Nahid Ahmadvand]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 18:16:23 GMT</pubDate>
            <atom:updated>2026-11-11T18:16:22.437Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Forecasting 48 Hours Ahead: How Much Complexity Was Worth It? An investigation of hourly ad views, the value of seasonal structure, and what a favorable accuracy ranking left unresolved." src="https://cdn-images-1.medium.com/max/1024/1*mJ6j_nDNnladdVpg_rdKYA.png" /></figure><h4><em>An investigation of hourly ad views, the value of seasonal structure, and what a favorable accuracy ranking left unresolved.</em></h4><p>At the end of an observed hour, I have a history of how many ads were watched in a mobile game. Ahead of me are 48 empty hourly slots. I need to fill all of them before another observation arrives.</p><p>The calendar is already known. I can identify tomorrow&#39;s hour and weekday, and I can look back at previous daily patterns. The next hour’s count is still unknown when I forecast the second hour, so I cannot use it as an observed predictor. A predictor that looks useful in a spreadsheet can quietly assume information that this forecasting task does not provide.</p><p>That constraint gave my investigation its starting point. Before comparing models, I needed to decide what each forecast could know and what would count as an improvement. Repeating the latest daily pattern offered a concrete reference. Any additional machinery would have to earn its place against that reference over the same future hours.</p><p>My central question was whether additional modeling complexity would improve this particular 48-hour task enough to justify it. Would a more detailed seasonal representation help? Would transforming the target, learning from lagged observations, or combining forecasts add something that a simpler procedure missed?</p><p>I worked through those questions using hourly advertising data from one Zeptolab game. The useful story emerged from the comparisons: which choices reduced error, which plausible ideas failed to help, and what happened when I asked the selected forecast to describe uncertainty as well as a predicted count.</p><blockquote>A note on presentation: this article covers the methodology, results, and interpretation, not a full code walkthrough. The complete notebook is linked at the end for anyone who wants the full implementation.</blockquote><h3>A forecast begins with its information set</h3><p>The dataset contains 2,085 hourly observations, from August 3, 2017, at 00:00 through October 28 at 20:00. The timestamps are complete and unique, and the counts are positive. No timezone was supplied, so I kept the dataset&#39;s clock without assigning it a geographical interpretation.</p><p>The data and initial lag-feature/PCA exercise came through <a href="https://mlcourse.ai/book/topic09/bonus_assignment09.html">mlcourse.ai&#39;s time-series material</a>. I extended that starting point into a comparison of complete forecasting procedures.</p><p>I began evaluation after 336 hours, or two weeks, of initial training. At each origin, a procedure used the observations then available, issued 48 predictions, and advanced by 48 hours. Refitting at the next origin allowed newly observed counts to enter the history.</p><p>This produced 29 non-overlapping development paths, covering August 17 through October 13. Each forecast path consisted of 48 consecutive hourly predictions issued from a single forecast origin. Six later paths covered October 14–25, providing 288 historical validation hours. A final 69-hour observed tail was excluded from evaluation targets and retained for the concluding refit.</p><p>The distinction between development and validation needs one qualification here: I had already examined the later outcomes during the broader research. Within the implemented run, selection uses development results before validation begins. That chronology is useful, but the later comparison remains historical validation, rather than an untouched confirmatory test. Likewise, exploratory analysis informed the specifications, so development is adaptive research.</p><p>I selected the procedure with the lowest mean per-origin <strong>Mean Absolute Scaled Error</strong>, or MASE. At each origin, I divided the next 48 hours&#39; mean absolute error by the average absolute 24-hour difference in that origin&#39;s training history:</p><figure><img alt="\operatorname{MASE}_i=
\frac{\frac{1}{48}\sum_{h=1}^{48}|y_{i,h}-\hat y_{i,h}|}
{\frac{1}{n_i-24}\sum_{t=25}^{n_i}|y_t-y_{t-24}|}." src="https://cdn-images-1.medium.com/max/443/6a8f59c69f7def20fbce66f78d92c8e25eae2d2a646028b1a147bbe91d32f8bf" /></figure><p>Here, n_i is the number of training observations available at origin i. The numerator measures the issued forecast&#39;s error. The denominator supplies a scale using only that training history. This helps compare origins at different traffic levels, while mean absolute error (MAE) in original count units remains useful for interpreting the size of misses. Root mean squared error (RMSE) adds a view that gives larger errors more weight. The forecasting text by Hyndman and Athanasopoulos provides <a href="https://otexts.com/fpp3/toolbox.html">background on these evaluation choices</a>.</p><p>One detail matters when reading the results: the MASE denominator is a historical seasonal-difference error, not the evaluated seasonal-naive forecast&#39;s future error. Consequently, the seasonal-naive benchmark need not score exactly one. A procedure can exceed one and still beat it on the actual target paths.</p><p>The common outer paths made the comparisons interpretable. Inside each training history, tuning also had to respect chronology. For feature-based models, that meant chronological cross-validation with 48-hour inner validation blocks, with scaling and learned component transformations fitted inside the tuning pipeline. The horizon shaped the training problem as well as the final prediction call.</p><p>The core evaluation logic was a chronological loop. This shortened excerpt uses the project&#39;s scoring helpers and omits result storage and prediction-validity checks:</p><pre>for _, r in DEV_FOLDS.iterrows():<br>    train = df.loc[df.index &lt; r[&quot;validation_start&quot;]]<br>    val_dates = pd.date_range(<br>        r[&quot;validation_start&quot;], r[&quot;validation_end&quot;], freq=&quot;h&quot;<br>    )<br>    preds = np.asarray(fit_predict(train, val_dates), dtype=float)<br>    actual = df.loc[val_dates, &quot;y&quot;].to_numpy()<br>    scores = point_scores(actual, preds, mase_scale(train))</pre><p>Here, fit_predict is the procedure being evaluated. The training slice ends before the target path, and mase_scale(train) recomputes the denominator from that history. Future actuals enter only when scoring the issued predictions.</p><h3>What the daily pattern explained and what it left open</h3><p>Before asking for a sophisticated model, I looked at the development series itself. The plot below shows why neither a constant level nor a fixed seasonal profile could be taken for granted.</p><figure><img alt="image.png" src="https://cdn-images-1.medium.com/max/1024/a1f537a1c6e2710b0a7ad45c44a2a18ec798f9d274556a1851f6c33974974b2b" /><figcaption>Figure 1. Hourly ad views during the development period, showing repeated daily cycles, rising traffic levels, and several sharp drops.</figcaption></figure><p>The mean of the first week in this region was about 42,664 ads per hour; the last week&#39;s mean was about 71,511. Superimposed on the changing level was a repeated daily shape. A forecast needed to follow both features without treating every local movement as a lasting trend.</p><p>I used a 24-hour seasonal-naive forecast as the main benchmark. It repeats the last observed day&#39;s profile across the entire 48-hour horizon. For the second forecast day, it continues repeating that known profile; it never substitutes an unobserved first-day actual. This made it a meaningful, information-feasible reference.</p><p>The calendar averages suggested another question. Was the daily cycle enough, or did a weekly representation deserve investigation?</p><figure><img alt="image.png" src="https://cdn-images-1.medium.com/max/1024/4c499c420eeb62622bcd336bc9e83e479dd894c3cfdbe6d8bcd3a4e2bb5dc419" /><figcaption>Figure 2. Average hourly ad views by hour of day and weekday during the development period. The hourly profile is lowest near 11:00 and highest near 21:00, while weekday averages also differ. These profiles summarize a period with a changing traffic level.</figcaption></figure><p>The highest hourly average was approximately 53% above the lowest; the weekday spread was about 19%. These are descriptive profiles. Changing levels and the observations represented in each group prevent treating them as isolated calendar effects. Still, they gave me a reason to examine more than one seasonal timescale.</p><p>Sharp drops raised a separate issue: should I alter the observations before fitting? A retrospective robust STL decomposition separated seasonal structure, trend, and a residual component. A residual diagnostic flagged seven hours on two dates. That identified unusual behavior, without establishing an outage or a measurement error. An interpolation sensitivity exercise changed the seasonal-naive result on one of the 29 paths. I retained <strong>raw observations for every primary forecasting comparison</strong>, keeping that exercise as a separate diagnostic.</p><p>Smoothing supplied the first model-based alternatives. Simple exponential smoothing updates a level; Holt adds a trend; additive Holt–Winters also updates a daily seasonal pattern. Their results showed why fitting an interpretable decomposition was not enough: Holt–Winters had mean development MASE 1.638, above seasonal naive&#39;s 1.393, and Holt was substantially worse.</p><p>The fitting procedures also differed. Level and trend models used one-step training squared error, while Holt–Winters was tuned on inner 48-hour blocks. Those are comparisons of the implemented procedures, with different tuning budgets. The evidence supported moving beyond these smoothing specifications, without establishing a general weakness of exponential smoothing.</p><h3>Giving weekly structure an explicit role</h3><p>Seasonal autoregressive integrated moving-average modeling, or SARIMA, offered a way to describe dependence between observations and past shocks, while using seasonal differencing to represent the daily pattern. A 24-hour difference replaces the level with the change from the same hour yesterday. Autoregressive terms describe dependence on earlier values of the modeled process; moving-average terms describe dependence on earlier innovations, or model shocks.</p><p>The stationarity diagnostics motivated candidates with no ordinary differencing and one daily seasonal difference. I searched a stated grid of 16 SARIMA specifications at each origin. The Akaike Information Criterion (AIC) balances fitted likelihood with a parameter-count penalty; a lower value is preferred. It selected among eligible fits using a common likelihood sample within each search. The outer 48-hour comparison then assessed forecasting performance. Finite estimates and convergence checks were required; those checks do not certify a global optimum. I recorded the fitting safeguards and diagnostic qualifications with the analysis.</p><p>Raw SARIMA reduced mean development MASE to 1.117. There was nevertheless a reason to investigate weekly structure: after daily differencing, the sample autocorrelation at lag 168 was 0.407. That correlation did not establish that daily SARIMA missed weekly dependence, since its dynamics can propagate across many days. It did give me a reason to test an explicit weekly representation.</p><p>Fourier terms provide a compact way to do that. Sine and cosine pairs describe a repeating curve, with additional harmonics allowing more detail. Their future values are known from the timestamps, so they introduce no dependence on unavailable future counts. This is the idea behind <a href="https://otexts.com/fpp3/dynamic-regression.html">dynamic harmonic regression</a>.</p><p>I constructed those predictors with this function:</p><pre>def fourier_terms(index, period, K, prefix):<br>    &quot;&quot;&quot;K sine/cosine pairs for one seasonal period, indexed by hours since epoch.&quot;&quot;&quot;<br>    t = (index.astype(&quot;int64&quot;) // 10**9) / 3600.0<br>    cols = {}<br>    for k in range(1, K + 1):<br>        cols[f&quot;{prefix}_sin{k}&quot;] = np.sin(2 * np.pi * k * t / period)<br>        cols[f&quot;{prefix}_cos{k}&quot;] = np.cos(2 * np.pi * k * t / period)<br>    return pd.DataFrame(cols, index=index)</pre><p>I compared two approaches. Dynamic harmonic regression represented both daily and weekly cycles with Fourier terms and used ARIMA errors without seasonal differencing. The hybrid retained daily seasonal differencing and added weekly Fourier regressors to <strong>SARIMA(1,0,0)(0,1,1)[24]</strong>.</p><p>In that hybrid, the nonseasonal AR(1) handles short-range dependence, while the seasonal MA(1) relates innovations separated by 24 hours. The Fourier regressors describe the weekly profile. At every origin, AIC chose the weekly harmonic count from K = 2, 3, or 4.</p><p>Mean development MASE was 1.048 for the full harmonic representation and 0.955 for the hybrid. The improvement made the hybrid a strong candidate, but it belonged to the complete specification and fitting rule. These alternatives changed more than one modeling choice, so the comparison did not isolate a universal benefit from adding Fourier columns.</p><p>That left a useful question: could changing the response scale improve on this seasonal modeling work?</p><h3>A plausible transformation that did not improve the comparison</h3><p>The development plots suggested that variability changed with traffic level. I therefore investigated Box–Cox transformations, estimating their parameter from the training history at each origin. A transformation can make some patterns easier to model, but forecasts ultimately had to return to the original count scale.</p><p>That return introduces a consequential distinction. A nonlinear inverse transformation generally does not turn a transformed-scale mean into an original-scale mean. Under the assumed Gaussian transformed predictive distribution, plain inversion gives a nominal median when the inverse is well defined. For the mean-corrected variant, I used a second-order approximation to the original-scale expectation.</p><p>The distinction appears directly in the back-transformation. Here, fitted is the transformed-scale SARIMA fit, and lam is the Box–Cox parameter estimated from training data:</p><pre>fc = fitted.get_forecast(steps=len(val_dates))<br>mu = np.asarray(fc.predicted_mean)<br>var = np.asarray(fc.var_pred_mean)<br>median = inv_boxcox(mu, lam)<br>mean_approx = median * (<br>    1 + var * (1 - lam) / (2 * (lam * mu + 1) ** 2)<br>)</pre><p>These are two summaries of the transformed model&#39;s forecast. The correction uses predictive variance; it does not refit the model to optimize original-scale absolute error.</p><p>The next figure shows both the fitted transformation parameter and the resulting forecast losses.</p><figure><img alt="image.png" src="https://cdn-images-1.medium.com/max/1024/a48f17e56171fd2855bad2a99fdb9dab0871733fd839978842a0a0d55ee3be36" /><figcaption>Figure 3. Box–Cox parameter estimates change across development origins (left). Transformed median and approximate-mean SARIMA forecasts are compared with untransformed SARIMA on the same paths using MASE (right). The horizontal red reference is an exploratory development-region estimate, not a value supplied to earlier forecasting fits. The plotted “bias-corrected” mean uses a second-order approximation.</figcaption></figure><p>Neither transformed version improved the full development comparison. Mean MASE was 1.236 for plain inversion and 1.194 for the approximate mean, versus 1.117 for raw SARIMA. Correcting the summary functional helped relative to plain inversion here, but did not make transforming the target beneficial overall.</p><p>I also checked whether carrying an order selected on the raw scale was disadvantaging the transformed model. On ten matched development paths, re-searching transformed orders gave MASE 1.286, compared with 1.295 for the fixed-order transformed procedure. Raw SARIMA scored 1.010 on those same paths.</p><p>That small transformed-order contrast changed sign across executions. The saved evidence does not identify its numerical cause, and the difference is too fragile to support a firm claim about which transformed-order strategy is better. Both remained worse than raw SARIMA on the matched targets.</p><p>There was also a distributional issue. A nonzero Box–Cox inverse has a restricted transformed domain; an unrestricted Gaussian predictive distribution can place mass outside it. Near a boundary, original-scale values can become extreme. A finite simulation or a Taylor correction does not resolve that support problem or establish a stable population mean.</p><p>The experiment therefore taught me more than “the transform lost.” It showed why choosing a response scale and choosing an original-scale forecast summary are connected decisions. For this task, neither supplied evidence strong enough to displace the raw-scale seasonal procedures.</p><h3>Learning from history without borrowing from the future</h3><p>Feature-based models approached the same task from another direction: represent historical counts and calendar information as predictors, then learn their relationship to the target. The crucial design choice was which lags could be used for all 48 forecast rows.</p><p>Let the last observed time be o. For forecast horizon h and lag ℓ, the source observation is at o + h − ℓ. It is available only if</p><figure><img alt="o+h-\ell\leq o \quad\Longleftrightarrow\quad h\leq\ell." src="https://cdn-images-1.medium.com/max/365/ae647ad139f3213443721f152434f730e5147b1612202d69c7a8d2a2381d1dad" /></figure><p>I therefore used lags 48 through 119. That is a block of 72 lagged predictors, ending at 48 + 72 − 1; 119 is the endpoint of the chosen feature window, not a separately optimized cutoff. Five deterministic calendar features brought the total to 77.</p><p>I built the lag and calendar features as follows:</p><pre>LAG_START, LAG_END = HORIZON, HORIZON + 72<br><br>def make_features(frame):<br>    x = frame[[&quot;y&quot;]].copy()<br>    for i in range(LAG_START, LAG_END):<br>        x[f&quot;lag_{i}&quot;] = x[&quot;y&quot;].shift(i)<br>    h, w = x.index.hour, x.index.dayofweek<br>    x[&quot;hour_sin&quot;], x[&quot;hour_cos&quot;] = np.sin(2*np.pi*h/24), np.cos(2*np.pi*h/24)<br>    x[&quot;weekday_sin&quot;], x[&quot;weekday_cos&quot;] = np.sin(2*np.pi*w/7), np.cos(2*np.pi*w/7)<br>    x[&quot;is_weekend&quot;] = (w &gt;= 5).astype(int)<br>    return x</pre><p>One fitted regressor predicts the 48 rows. It does not feed its earlier predictions back as if they were observations. The preparation code leaves future targets missing and asserts that every lag points to an observation available at the origin, including in inner validation.</p><p>With many related lag columns, regularization was a natural comparison. Ridge constrains coefficient size; Lasso can also set coefficients to zero. For Lasso, the essential tuning logic combined scaling and regression inside one pipeline:</p><pre>est = Pipeline([<br>    (&quot;s&quot;, StandardScaler()),<br>    (&quot;m&quot;, Lasso(max_iter=200000, tol=1e-3)),<br>])<br>search = GridSearchCV(<br>    est, grid, cv=forecast_cv(X),<br>    scoring=&quot;neg_mean_absolute_error&quot;,<br>).fit(X, y)</pre><p>Here, grid is the declared Lasso penalty grid, and forecast_cv(X) supplies chronological 48-hour validation blocks. Scaling is learned separately inside each inner training fit. MAE selects the penalty within an origin; mean development MASE compares the complete procedures across origins.</p><p>Mean development MASE fell from 1.344 for ordinary linear regression to 1.331 for Ridge and 1.277 for Lasso.</p><p>Principal component analysis (PCA) and partial least squares (PLS) asked whether a smaller representation would help. PCA preserves predictor variance; PLS uses the response when constructing components. Neither objective guarantees the best forecast. PCA&#39;s fixed 95%-variance rule scored 1.335, its cross-validation-selected component count 1.372, and PLS 1.370. Their aggregate ordering did not identify why tuning or a particular representation helped or hurt.</p><p>The fitted-feature plots provided a closer look at what these representations used.</p><figure><img alt="image.png" src="https://cdn-images-1.medium.com/max/1024/534a6a7f9de00f133e81015df67cf2152a0e052d46119be46ed4b3617f3c0d38" /><figcaption>Figure 4. Three panels show Lasso coefficients, XGBoost gain importances, and PCA loadings across lag and calendar predictors. These come from a separately tuned feature-inspection probe at the last development origin. Coefficient magnitude, tree gain, and PCA loading describe different aspects of these particular fits.</figcaption></figure><p>The Lasso probe retained 55 of 77 predictors, while the 95%-variance PCA probe retained eight components. Those are properties of those fits, not evidence that every origin selected the same features. Correlated lags can share information with calendar variables, so I did not read their apparent importance as separate causal contributions.</p><p>I tested recency directly by comparing expanding training histories with three- and six-week target windows at the same origins. Three weeks improved Lasso from 1.277 to 1.248, while six weeks gave 1.348. Ordinary linear regression worsened under both windows. “Use recent data” was therefore not a shared rule across estimators.</p><p>Chronological target encoding gave another modest improvement. Training rows used strictly earlier hour/weekday group averages, and forecast encodings were frozen at issuance. The encoded linear procedure scored 1.284. However, requiring prior group history also removed early training rows. Its gain belongs to the combined representation-and-sample procedure, rather than a clean encoding-only effect.</p><p>Finally, I compared XGBoost objectives. Absolute-error-based evaluation makes an absolute-error training loss worth trying: in an unrestricted population problem, a conditional median minimizes expected absolute loss. But a regularized finite tree fit only estimates that target. Squared-error XGBoost scored 1.240; absolute-error XGBoost scored 1.331. Objective alignment supplied a reason to experiment, not a guarantee that the fitted procedure would forecast better.</p><h3>Would a different seasonal model or a combination add something?</h3><p>Prophet offered a further representation of changing trend and seasonal profiles. I compared additive seasonality with multiplicative seasonality, where seasonal amplitude varies with the modeled trend. The multiplicative procedure improved on the additive one, with development MASE 1.201 versus 1.233.</p><p>I also investigated Box–Cox Prophet variants summarized by an original-scale simulated median or mean. Both performed worse: 1.643 and 1.784, respectively. These summaries came from separate 1,000-draw runs conditional on fitted parameters, rather than full Bayesian parameter uncertainty. Omitting undefined inverse draws also conditions the result on the retained draws. The negative result is informative, while those implementation details limit an interpretation based solely on mean-versus-median theory.</p><p>Having obtained several reasonable forecasts, I asked whether combining them could improve on selecting one. I used a fixed pool of ten named configurations and compared equal weights, a component-wise median, and inverse-MAE weights. Some families contributed multiple configurations, so this was not a pool of ten independent sources of information.</p><p>The performance-based weights used strictly earlier development losses, starting with equal weights. Validation and final weights used development losses only. This kept future outcomes out of each issued weighting decision, while pool design remained part of the adaptive investigation.</p><p>Averaging can help when members make partially offsetting errors, but their error magnitudes and biases matter as well. Pooled pairwise error correlations averaged 0.748; different model names did not imply independent errors. The correlations alone could not establish whether a combination would improve on the strongest member.</p><p>All three combination rules beat seasonal naive. Inverse-MAE weighting was the best combination at 1.098, followed by equal weights at 1.116 and the median at 1.125. None beat the weekly-Fourier hybrid. Equal weights remained the designated ensemble reference for the later shortlist; that choice was not a claim that it had won the combination comparison.</p><p>At this point, the cross-family results could be read together. The table retains a compact selection from the 32 procedures, all evaluated on the same 29 development paths.</p><figure><img alt="Screenshot 1405-08-19 at 19.18.57.png" src="https://cdn-images-1.medium.com/max/1024/589fc5327cdeec415e7e5303bc1215f321dae8c573f3b9ce779263402dc921e9" /><figcaption>Table 1. Selected development results, rounded to three decimals. These are complete forecasting procedures with their stated fitting and tuning rules; the full analysis reports all 32.</figcaption></figure><p>The most useful distinction was between kinds of complexity. Explicit weekly structure helped the leading procedure. More flexible learners, transformations, and combinations offered alternatives without consistently reducing the selected loss. That was a result about this series, horizon, and set of procedures, rather than a contest between classical statistics and machine learning.</p><h3>Looking beyond the first row</h3><p>The weekly-Fourier hybrid&#39;s mean development MASE was 0.954947 at the saved table&#39;s precision. It beat seasonal naive on 25 of 29 paths, with a mean paired difference of approximately −0.438. That gave the observed improvement more substance than a first-place rank alone.</p><p>I also examined paired loss differences with dependence-aware references. A heteroskedasticity- and autocorrelation-consistent (HAC) variance estimate allows for changing variance and serial correlation in the losses. The reported HAC comparison had p &lt; 0.0001, and a circular-block bootstrap gave a nominal 95% basic interval of [−0.558, −0.306]. Multiplicity adjustments accompanied the benchmark comparisons. A nominal 90% <a href="https://www.kevinsheppard.com/files/teaching/mfe/advanced-econometrics/Hansen_Lunde_Nason.pdf">Model Confidence Set</a>, which eliminates candidates using loss comparisons, retained the hybrid and the full harmonic-regression procedure.</p><p>These calculations organize evidence under assumptions about the loss process. They do not erase adaptive specification choices, establish independence between paths, or give a 90% probability that a surviving model is best. Non-overlapping target periods can still share temporal dependence and overlapping training histories.</p><p>I wanted to see how much the picture changed with the occasions evaluated. The following diagnostic brings together horizon errors, chronological performance, fold variation, and a reuse of the saved losses in small subsets.</p><figure><img alt="image.png" src="https://cdn-images-1.medium.com/max/1024/4866485d2c633cca507766b553ac5f2cb092b0d294acd854bd5066faa9ca421e" /><figcaption>Figure 5. Four diagnostic panels show error by forecast horizon, chronological performance as training history grows, fold-level losses, and first-place frequencies in random five-path subsets. The subset panel reuses existing development losses without refitting; the history-size panel also changes calendar period.</figcaption></figure><p>Across 3,000 random five-path subsets, the full-sample-selected leader ranked first 49% of the time, while 19 procedures took first place somewhere. These are overlapping, reused subsets, not new experiments or probabilities of winning on future data. They illustrate how a small target set can change the story.</p><p>The leader&#39;s mean absolute error was about 2,963 in the first 12 hours and 4,697 in the last 12, with a non-monotone curve between them. Meanwhile, the plot against training-history size also moved through calendar time. I read it as a performance trajectory, rather than attributing improvement entirely to more observations. The matched-origin rolling-window experiments addressed a narrower question about training recency.</p><h3>The later period separated accuracy from reliability</h3><p>For historical validation, I refitted the shortlisted procedures at each of the six later origins. Earlier validation observations could enter subsequent training histories; selection rules and combination weights remained development-derived. Four candidates entered the reported ranking:</p><figure><img alt="Screenshot 1405-08-19 at 19.47.16.png" src="https://cdn-images-1.medium.com/max/1024/84fe59d299103929f69b8308e695dbdabaad4544d56b9315be6ad11b7e908156" /><figcaption>Table 2. Historical validation on six 48-hour paths. MASE retains the saved three-decimal precision; MAE is rounded to whole ad counts.</figcaption></figure><p>The hybrid led again and beat seasonal naive on four paths. Its paired-t p-value was 0.3150, with Holm-adjusted p = 0.9449. The observed reduction remained useful descriptively, while six potentially dependent paths and prior exposure limited the inferential claim. Non-significance established neither equivalence nor a particular explanation for the uncertainty.</p><p>The path plots added information that the averages could not show. Here I focused on where the selected forecasts and their intervals missed the actual observations.</p><figure><img alt="image.png" src="https://cdn-images-1.medium.com/max/1024/642071cb2e02b660bf69d32066710ab934868330ab7b6f9bffb810ff083e7ed1" /><figcaption>Figure 6. Six historical validation paths compare actual ad counts with weekly-Fourier, equal-weight combination, and seasonal-naive forecasts. Shading shows the weekly-Fourier procedure&#39;s nominal 80% pointwise interval, with several actual peaks exceeding its upper limit. The full harmonic-regression candidate is reported in Table 2 but is not plotted here.</figcaption></figure><p>Several actual peaks exceeded the forecast and upper limit. I treated this apparent imbalance as a visual observation, since I hadn’t saved separate counts of misses above and below the intervals. I couldn’t establish its cause from the plots alone. The coverage results gave me a firmer conclusion: the procedure with the lowest average point error still fell short of its nominal interval coverage.</p><h3>What did an “80% interval” mean here?</h3><p>I used earlier forecast errors to construct two uncertainty summaries. <strong>Pointwise intervals</strong> address coverage at each individual horizon across forecast occasions. <strong>Simultaneous bands</strong> address whether all 48 observations fall inside their bands on the same occasion. Covering 47 hours and missing one contributes 47 hourly successes, but fails whole-path coverage.</p><p>For pointwise intervals, I took an order statistic of absolute historical errors separately at each horizon and placed that distance above and below the new point forecast. For simultaneous bands, I instead calibrated a maximum standardized error across the whole path, using horizon-specific scales estimated from the first eight development paths.</p><p>Both constructions borrowed the conformal calibration rank:</p><figure><img alt="k=\left\lceil(m+1)(1-\alpha)\right\rceil," src="https://cdn-images-1.medium.com/max/290/71c2ab89f5e5591435a1e6a19684039bf763fd95d3e24b5ed4fab20db85c08cb" /></figure><p>where m is the number of calibration paths and 1 − α is nominal coverage. If the required rank exceeded the available scores, the implementation returned an unbounded band and marked it unsupported for finite-band evaluation. It did not clip the rank to manufacture support.</p><p>The threshold calculation follows that rule directly:</p><pre>def conformal_quantile(scores, alpha):<br>    scores = np.asarray(scores, dtype=float)<br>    m = len(scores)<br>    k = int(np.ceil((m + 1) * (1 - alpha)))<br>    return np.inf if k &gt; m else float(np.partition(scores, k - 1)[k - 1])</pre><p>For pointwise intervals, I applied this function separately at each horizon to one absolute residual per calibration path. The resulting threshold sets the distance above and below the new forecast; the 48 errors within a path are not pooled as independent calibration examples.</p><p>This correction has a role in <a href="https://www.stat.cmu.edu/~ryantibs/papers/conformal.pdf">conformal prediction</a>, but chronological separation alone does not establish the exchangeability needed for its usual coverage argument. Dependence, changing traffic, and refitting leave a distribution-free guarantee unestablished here. I evaluated these intervals empirically.</p><p>During validation, calibration was frozen using 21 development paths after the eight scale-estimation paths. No validation residuals entered the thresholds or the predictive residual distribution, even though point models continued to refit. The resulting comparison was:</p><figure><img alt="Screenshot 1405-08-19 at 19.51.15.png" src="https://cdn-images-1.medium.com/max/1024/d0c5b10143caecf240367a167bb0705cb48cfd916f20874192d76ceca0888bff" /><figcaption>Table 3. Coverage, width, and proper scores describe different aspects of uncertainty. Lower interval score and CRPS are better. CRPS evaluates a separate signed-residual predictive distribution, not the absolute-residual intervals.</figcaption></figure><p>The combination covered more hours, but its intervals were wider and its interval score was worse. Interval score penalizes both width and misses, preventing coverage alone from rewarding indiscriminately broad bands. The hybrid had the better score while still falling short of nominal coverage. Neither observation cancels the other.</p><p>The continuous ranked probability score (CRPS) evaluated another issued object: an empirical predictive distribution made by adding signed development residuals, actual minus forecast, at each horizon to the new forecast. It assesses a full distribution, unlike an interval score for a specified pair of bounds. The <a href="https://sites.stat.washington.edu/raftery/Research/PDF/Gneiting2007jasa.pdf">proper-scoring literature</a> explains why those distinctions matter when evaluating probabilistic forecasts.</p><p>There were six forecast paths, not 288 independent coverage trials. I did not calculate uncertainty intervals for these coverage or score contrasts. Separate, wider simultaneous bands contained all six paths, but that addressed a different target and did not repair pointwise undercoverage.</p><p>This changed what I could conclude from the favorable ranking. I had a promising point-forecasting procedure on the observed periods, alongside an uncertainty method that needed further work. The comparison did not isolate calibration drift or another single cause of failure. It did establish a practical weakness: the selected procedure&#39;s nominal 80% pointwise interval covered only 70.83% of these validation hours.</p><h3>Returning to the 48 empty slots</h3><p>I finished by refitting the selected SARIMA-plus-weekly-Fourier rule on all 2,085 observations, including the tail excluded from evaluation. The final AIC comparison selected <strong>K = 4</strong>, giving four sine/cosine pairs, or eight weekly Fourier regressors.</p><p>The final origin was October 28, 2017, at 20:00. The issued path covered 21:00 that evening through October 30 at 20:00.</p><figure><img alt="image.png" src="https://cdn-images-1.medium.com/max/1024/c4a1762d43449641613eb03a4d80f4efba25d7ac5cec61f35568314ccefc1c4c" /><figcaption>Figure 7. Observed history ends on October 28, 2017, at 20:00, followed by 48 weekly-Fourier predictions, an empirical nominal 80% pointwise band, and a dashed seasonal-naive forecast. The band labeled “rolling-residual” uses frozen development residuals for this forecast; it does not update with future actuals.</figcaption></figure><p>The predictions ranged from approximately 80,180 to 126,393 ads per hour. Seasonal naive lay inside the pointwise band on 24 of 48 steps. That is containment of one forecast by another forecast&#39;s interval, not actual coverage. Future observations are unavailable for this path.</p><p>The final issue time also introduced a transfer assumption. Development calibration and validation paths began at midnight after 23:00 origins. This path began at 21:00 after a 20:00 origin. A residual at horizon one now corresponds to a different clock hour. I retained the development calibration without having evaluated that schedule change, so the band carries both the observed undercoverage and an additional untested transfer.</p><p>The most useful next experiment would therefore evaluate a fixed procedure on a previously unexamined period, at the intended issue schedule. I would assess both point accuracy and uncertainty, preserving which observations were available and when calibration could update. Directional misses and alternatives to symmetric error bands deserve attention alongside any new model family.</p><p>The current study remains limited by one short series, adaptive development, prior validation exposure, and dependent losses. Numerical replay also needs care: the small Box–Cox contrast changed across executions, and a clean installation has not been tested. The <a href="https://github.com/Nahid-ahmdv/hourly-ad-view-forecasting/blob/main/README.md">project README</a> records the environment and data-access qualifications.</p><p>Additional complexity earned its place selectively. Weekly seasonal structure contributed to the leading complete procedure; transformations, supervised representations, and combinations did not consistently improve on it. The result gave me a specific method to build on, rather than a general rule that simpler or classical models always win.</p><p>Returning to those 48 empty slots, I could now explain why I chose the forecast that filled them. I could also explain what its shaded band had not established. Following the investigation through that second question made the result more useful than stopping at the lowest error.</p><h3>Sources and companion material</h3><ul><li><strong>Research record:</strong> <a href="https://github.com/Nahid-ahmdv/hourly-ad-view-forecasting/blob/main/hourly_ads_forecasting_final.ipynb">complete project implementation and saved results</a> and <a href="https://github.com/Nahid-ahmdv/hourly-ad-view-forecasting/blob/main/README.md">project README</a>. All empirical results and figures above come from the saved analysis outputs; no forecasting experiments were rerun for this article.</li><li><strong>Data and educational starting points:</strong> Yury Kashnitsky, <a href="https://mlcourse.ai/book/topic09/bonus_assignment09.html">mlcourse.ai Bonus Assignment 9</a>; Dmitriy Sergeyev, <a href="https://mlcourse.ai/book/topic09/topic9_part1_time_series_python.html">Time series analysis in Python</a>; Egor Polusmak, <a href="https://mlcourse.ai/book/topic09/topic9_part2_facebook_prophet.html">Forecasting with Prophet</a>. The Zeptolab data and initial lag/PCA setup came through these course materials. The two public course articles identify their license as <a href="https://creativecommons.org/licenses/by-nc-sa/4.0/">CC BY-NC-SA 4.0</a>; the bonus assignment has separate terms. This article grants no new rights over third-party material.</li><li><strong>Forecast design and seasonal modeling:</strong> Hyndman, R. J., and Athanasopoulos, G., <em>Forecasting: Principles and Practice</em>, third edition: <a href="https://otexts.com/fpp3/toolbox.html">the forecaster&#39;s toolbox</a> and <a href="https://otexts.com/fpp3/dynamic-regression.html">dynamic regression models</a>.</li><li><strong>Implementation references:</strong> <a href="https://www.statsmodels.org/stable/generated/statsmodels.tsa.statespace.sarimax.SARIMAX.html">statsmodels SARIMAX</a> and <a href="https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.TimeSeriesSplit.html">scikit-learn TimeSeriesSplit</a>. The project records the versions used in the analysis; current documentation is linked for orientation.</li><li><strong>Model comparison:</strong> Hansen, P. R., Lunde, A., and Nason, J. M. (2011), <a href="https://www.kevinsheppard.com/files/teaching/mfe/advanced-econometrics/Hansen_Lunde_Nason.pdf">The Model Confidence Set</a>.</li><li><strong>Calibration and scoring:</strong> Lei, J., G&#39;Sell, M., Rinaldo, A., Tibshirani, R. J., and Wasserman, L. (2018), <a href="https://www.stat.cmu.edu/~ryantibs/papers/conformal.pdf">Distribution-Free Predictive Inference for Regression</a>; Gneiting, T., and Raftery, A. E. (2007), <a href="https://sites.stat.washington.edu/raftery/Research/PDF/Gneiting2007jasa.pdf">Strictly Proper Scoring Rules, Prediction, and Estimation</a>. These sources explain methodology; their assumptions are not established merely by using related constructions in a time series.</li></ul><p>Thanks for reading. Questions, corrections, and ideas for extending the analysis are welcome in the comments.</p><h4>Links</h4><ul><li>Full notebook + code: <a href="https://github.com/Nahid-ahmdv/hourly-ad-view-forecasting">View the complete implementation</a></li><li>LinkedIn: <a href="https://www.linkedin.com/in/nahid-ahmadvand-00382b365/">Nahid Ahmadvand</a></li><li>Website: <a href="https://www.nahidahmadvand.website/">nahidahmadvand.website</a></li></ul><p><em>This story was drafted on </em><a href="https://tk.xyz"><em>TK</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cf4308edce3b" width="1" height="1" alt=""><hr><p><a href="https://thequantasticjournal.com/forecasting-48-hours-ahead-how-much-complexity-was-worth-it-cf4308edce3b">Forecasting 48 Hours Ahead: How Much Complexity Was Worth It?</a> was originally published in <a href="https://thequantasticjournal.com">The Quantastic Journal</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Plug That Outlived the Bottle]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/the-plug-that-outlived-the-bottle-5d9027d41d79?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/1672/1*8-uYe7z3H3sMZYDXoNMcjA.png" width="1672"></a></p><p class="medium-feed-snippet">Why Every Wine Cork Breathes Differently</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/the-plug-that-outlived-the-bottle-5d9027d41d79?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/the-plug-that-outlived-the-bottle-5d9027d41d79?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/5d9027d41d79</guid>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[wine]]></category>
            <category><![CDATA[chemistry]]></category>
            <dc:creator><![CDATA[shanmugam mukundan]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 15:57:48 GMT</pubDate>
            <atom:updated>2026-11-11T15:57:45.540Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Hidden Ways Insects Support Animal Life — From Habitat To Health — Go Way Beyond Being Food]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://thequantasticjournal.com/the-hidden-ways-insects-support-animal-life-from-habitat-to-health-go-way-beyond-being-food-20445ac5e97f?source=rss----cf522ff06849---4"><img src="https://cdn-images-1.medium.com/max/1667/1*AYZn510JuuppdDExyxRViw.png" width="1667"></a></p><p class="medium-feed-snippet">As insect biodiversity declines, people and all the other species on Earth stand to lose the beneficial &#x2018;ecosystem services&#x2019; bugs provide.</p><p class="medium-feed-link"><a href="https://thequantasticjournal.com/the-hidden-ways-insects-support-animal-life-from-habitat-to-health-go-way-beyond-being-food-20445ac5e97f?source=rss----cf522ff06849---4">Continue reading on The Quantastic Journal »</a></p></div>]]></description>
            <link>https://thequantasticjournal.com/the-hidden-ways-insects-support-animal-life-from-habitat-to-health-go-way-beyond-being-food-20445ac5e97f?source=rss----cf522ff06849---4</link>
            <guid isPermaLink="false">https://medium.com/p/20445ac5e97f</guid>
            <category><![CDATA[science]]></category>
            <category><![CDATA[nature]]></category>
            <category><![CDATA[entomology]]></category>
            <category><![CDATA[animals]]></category>
            <category><![CDATA[ecosystem]]></category>
            <dc:creator><![CDATA[The Conversation U.S.]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 18:08:45 GMT</pubDate>
            <atom:updated>2026-11-10T18:08:44.311Z</atom:updated>
            <cc:license>https://creativecommons.org/licenses/by-nd/4.0/</cc:license>
        </item>
    </channel>
</rss>