Alexander Anikin

Spent some time thinking through a question: "Should we just let an LLM do the MMM/attribution work now?" Short answer: No. Longer answer: There are three places LLMs are genuinely useful in marketing analytics right now, and at least four places where the hype is actively dangerous if you're not careful about it. Wrote up where I think the line actually sits. Where we actually stand today, and where I think the real engineering effort is worth going next! https://lnkd.in/gdRzek_A

48

3 комментария

Online time series forecasting must handle concept drift, but the problem with most approaches is that they treat drift as a one-off event. This work redefines drift into Recurring Drift (reappearing patterns) and Emergent Drift (novel patterns). It proposesDynaME: a hybrid framework that selects specialized experts for recurring patterns while relying on a stable general expert under high-uncertainty scenarios. Experiments show that the method adapts effectively to both drift types and outperforms existing methods.

1

1 комментарий

Everyone is looking for the best forecasting model. That’s usually the wrong question. Building forecasting systems for unstable economies: case of Russia (2/12) In the previous post we ended up with a rather uncomfortable setup: • ~60 observations • reconstructed monthly data • structural breaks • changing regulations At this point the natural question is: which model should we use? This is where most discussions start. And in my experience, this is where things usually go in the wrong direction. Because with short and unstable time series, no single model will consistently work. Sometimes ARIMA performs better. Sometimes gradient boosting. Sometimes nothing works particularly well. So instead of trying to find the “best” model, I switched to a different approach: build an architecture. The first layer is a baseline model. For me that role is played by ARIMA. It captures the core dynamics of the time series and gives a reference forecast. Not perfect — but stable. The next layer includes models that can incorporate external factors. First — linear models: Linear Regression Ridge Lasso ElasticNet Then machine learning: CatBoost XGBoost LightGBM Random Forest Gradient Boosting Each model is tested with different lag structures. Then comes backtesting. And finally, the forecasts are combined using model blending, where weights depend on performance. This solves two problems: • short time series • nonlinear relationships But almost immediately another issue appears. Exogenous variables do not behave nicely. Some have 15 years of history. Some have five. Some start exactly when you need them the most. So the next question becomes: how do you select and structure exogenous variables when both data and economics are unstable? That’s where things stop being purely technical. Next post: economics vs machine learning in selecting exogenous variables.

5

6 комментариев

Most data teams spent the last decade optimizing speed: Time-to-Dashboard, query latency, time-to-insight. However, users don’t actually need faster SQL, but a correct answer they can confidently act on quickly. Time-to-Trusted-Answer (TTTA) is a KPI that correlates time and correctness by measuring the full journey from question to correct, explainable and consistent answer, and turns “trust” from a vague goal into an operational standard. I wrote a deep dive here (with examples, tooling, and an implementation blueprint): https://lnkd.in/gmwzJB94 Curious: what’s the biggest trust blocker in your org today—definitions, freshness, lineage, or evaluation?

17

Since neural networks rely on gradients, they are usually biased to smooth solutions. This can make NNs more effective for sequential data. On the other hand, tree-based models are more promosing for tabular data. This article, offers useful insights for anyone choosing between tree-based models and neural networks. https://lnkd.in/dbJrZiex #Machine_Learning #Tabular_Data #Tree_Based_Moeels #Neural_Network

7

Over the last few months, I've been working intensively with AI agents for enterprise use cases, either supervising or developing them myself. Through this experience Ive kept noticing a sometimes frustrating pattern: recurring mistakes still show up whenever a task requires working with exact elements within a set. For example: - counting occurrences - detecting duplicates - comparing sets - checking coverage - identifying what's missing or extra in a list The classic example is: "How many r's are in strawberry?" While simple, in enterprise workflows, similar situations appear all the time: duplicated owners, missing checks, inconsistent lists, simple coverage validations. These are not particularly complex problems, but they are not semantic problems; they are symbolic problems. Modern models have improved significantly at handling these tasks. Techniques such as Chain-of-Thought reduce the error rate considerably in many scenarios. However, Chain-of-Thought is still a probabilistic reasoning strategy. It can improve an answer, but it does not make the validation itself deterministic. On the other hand, tools like Python and SQL can solve these problems perfectly, but they are general-purpose tools that we typically associate with larger analyses or more complex workflows. That led me to a question: Is there a lightweight layer in between? Something simpler than Python, SQL, Prolog, or formal reasoning systems, but more reliable than asking the model to reason internally under probabilistic terms. As an experiment, I built a small proof of concept called Lightweight Symbolic Validator It's an experimental MCP tool that allows an agent to delegate symbolic checks to deterministic code whenever an exact answer is required. The goal is to reduce the need for "guessing" in situations where the answer should be fully verifiable. After testing it, my conclusion is more nuanced than I initially expected. Current frontier models already solve many of these cases correctly, but don't provide any deterministic guarantee. The tool, under certain scenarios, generates a significant improvement on the response quality vs the no-tool scenario, although it requires further tuning and testing to be effective in a greater scale. Still, the experiment left me with an interesting question: Should agents have a lightweight symbolic validation layer available by default? I'm talking about simple operations over: sets, sequences, duplicates coverage, basic relations I've published the experiment on GitHub for anyone who wants to try it, provide feedback, or challenge the idea: https://lnkd.in/dpsSfbNQ More than a finished solution, I see it as an exploration of a possible middle layer between semantic reasoning and heavier symbolic systems. Happy to hear from people experimenting with agents development

28

2 комментария

4 Agentic AI Patterns in Pricing Systems In markets with high volatility, pricing needs more than static rules. At our platform, we see Agentic AI pricing systems commonly follow four practical patterns. 1. Reflection Pattern The AI reviews and improves its own pricing decision. Example: After proposing a new price for a volatile raw material, the system rechecks recent demand response, margin impact, and past outcomes before finalizing. 2. Tool Use Pattern The AI actively uses external data sources. Example: The system pulls market indices, supplier cost updates, competitor prices, and inventory levels before recommending a price change. 3. Planning Pattern The AI plans multiple steps before acting. Example: Instead of a sudden price hike, the system plans a phased adjustment over multiple periods to reduce customer shock and manage volume risk. 4. Multi-Agent Pattern Multiple AI agents collaborate on pricing. Example: One agent tracks market volatility, another forecasts demand, and a policy agent ensures margin and compliance rules are met before a price is published. Why this matters Volatile markets demand pricing systems that can think, plan, and adapt. Agentic AI enables structured, governed, and faster pricing decisions at scale. #AgenticAI #PricingSystems #AIinPricing #DynamicPricing

7