My Tracking Notes: Building a Backtest Spreadsheet for AI Lottery Predictors

It was one rainy Wednesday night in Charlotte this past February when I realized I had gone too far. The blue light of my monitor was the only thing illuminating the home office, and the rhythmic, mechanical click of my keyboard at midnight while the rest of the house was completely silent had become my new soundtrack. From the hallway, my wife asked if I was finally coming to bed. I told her I just needed to input one more Saturday draw into the master sheet. She didn't buy it, and honestly, I didn't either.

I’m a data analyst by trade, so when my coworkers started an office Powerball pool, the 'birthday strategy' they used—picking numbers based on kids' ages and anniversaries—hit me like a physical weight. It felt messy. It felt inefficient. I started thinking about how 'random' feels like a personal challenge to a data analyst, even when the math says otherwise. That challenge led me down a rabbit hole of neural networks and AI lottery tools. I wanted to see if these platforms actually did anything better than a random number generator, or if they were just fancy-looking randomizers themselves.

The Setup: Turning the Powerball into a Dataset

To really test these tools, I couldn't just buy a ticket and hope. I needed a way to backtest their 'Top 10' or 'Hot' suggestions against what actually happened. I started with the basics of the game: the Powerball white ball pool consists of 69 numbers, and the red ball pool has 26. The odds of hitting that 5+1 combination are roughly 1 in 292,201,338. Those are the kind of numbers that make most people shrug and pick their lucky 7, but for me, it was the baseline for a spreadsheet.

I signed up for three different AI prediction platforms that claimed to use machine learning—specifically Long Short-Term Memory (LSTM) models—to find patterns in historical draws. My goal wasn't to 'hack' the system, but to document whether these tools could consistently outperform a truly random pick over a significant period. I set up my master sheet to track every Monday, Wednesday, and Saturday draw, which is the standard Powerball frequency. I've been maintaining this since late winter 2026, and the data has been... enlightening, to say the least.

The core of the process was building a VLOOKUP system. On one tab, I’d paste the AI's suggested number sets for the upcoming draw. On the second tab, I’d pull the actual winning numbers from the official 10:59 p.m. ET draws. A third tab calculated the 'hit rate'—how many numbers the AI actually got right compared to a control group of random numbers I generated myself using a simple Excel formula.

The Process: Tracking the Neural Networks

By early spring, my routine was set. Every Tuesday, Thursday, and Sunday morning, I’d grab my coffee and update the sheet. I wasn't just looking for the jackpot; I was looking for any statistical deviation. If an AI tool suggests 10 numbers and three of them hit consistently, that’s a pattern. If it suggests 10 and we get zero hits for three weeks, that’s just noise.

I should probably mention that I have zero formal training in professional gambling or lottery management. I’m just a guy who likes clean data. I’m not a math professor, and I’m certainly not an insider. If you find yourself spending money you don't have on these games, please stop and talk to a professional counselor. This is a hobby with a negative expected value, and I treat it as such—it’s entertainment, not a retirement plan.

I wrote about the early days of this project in my post about the spreadsheet that annoyed my wife, where I detailed the specific Excel formulas I used to automate the comparison. What I found as the months went by was that the AI tools often suffered from 'over-fitting.' They were so focused on the historical frequency of certain numbers that they often missed the 'cold' numbers that are just as likely to show up in a truly random draw. The tools were essentially trying to find a narrative in a series of independent events.

The Turning Point: Shifting the Strategy

After the first three months of tracking, the data was telling a very specific story. The hit rates for the AI-predicted numbers weren't significantly better than my random control group when it came to the white balls. However, I noticed something interesting about the payout distributions. Most people—and most basic AI tools—focus on the jackpot. But the jackpot odds are so astronomical that the data is basically a flat line of failure.

I started looking at the secondary prize tiers. The real edge isn't in predicting the 1-in-292,201,338 jackpot; it's in spotting where the payout distribution for secondary tiers gets weird. In some draws, the number of winners in the lower tiers is significantly lower than expected, meaning if you do hit those numbers, you aren't sharing the prize pool with as many people. I began tweaking my spreadsheet to track not just hits, but 'prize density.' I wanted to know if the AI could help me pick numbers that were less likely to be picked by the general public, thereby increasing the value of a non-jackpot win.

This contrarian approach felt more 'data-analyst' to me. Instead of trying to predict the future (impossible), I was trying to exploit human behavior (very possible). Humans love patterns—they love sequences, multiples of five, and dates. If an AI tool can help you lean into the 'ugly' numbers that others avoid, that’s where the actual utility lies. My six-month spreadsheet obsession has shown me that while the 'prediction' aspect is mostly a wash, the 'selection' aspect is where the math gets interesting.

Reflections from the Mid-July Audit

By mid-July, I sat down for a full audit of the past six months. I had hundreds of draws recorded and thousands of AI suggestions logged. The results were clear: the AI tools weren't magic. They didn't have a crystal ball. But they did provide a structured way to play that removed the emotional 'birthday' bias. For a data guy, that’s worth something, even if it’s just the peace of mind that comes with a structured process.

I’ve learned that the lottery is the ultimate stress test for any data model. It’s pure, unadulterated randomness. Even the most sophisticated LSTM model is just making an educated guess based on a very small sample size of historical draws. My wife still thinks the tracking is overkill, and as I look at the 50th version of my 'Backtest_Master_V2_Final_FINAL.xlsx' file, I realize she’s probably right. But there’s a certain clarity in seeing the math for what it is. It turns a game of desperate hope into a series of data points, and for me, that makes the whole thing a lot more tolerable.

If you’re going to try this, do it for the data, not the dollars. Keep your expectations grounded in reality. Past returns in any data set don't predict future ones, and that's especially true when you're dealing with gravity-fed plastic balls in a drum. Treat it like a puzzle, keep your spreadsheet clean, and remember that at the end of the day, the house always has the better algorithm.