The essence of our dynamic pricing model is to take the burden of pricing off your shoulders and make decisions for you that, as a human, you can no longer make precisely, consistently and well in the long term.
The dynamic pricing model is a mathematical system that:
regularly (e.g. hourly, every 6 hours, daily) optimizes the prices of your products,
for each product, it looks for the price point where it generates the most profit in a stable way,
and adjusts prices so that the customer encounters fair, understandable, optimal prices.
It does not promise infinite profit. The outcome is still influenced by:
how good your unit costs are,
how well thought-out your store’s operating costs are,
how good your marketing and traffic are.
What we can promise is this: we mathematically get the most out of the given situation,
consistently, without human error and guesswork.
Imagine that the prices of three of your products look like this:
What do we see from this?
Product 1: we slightly lower the price → cheaper, more customers, overall more profit.
Product 2: we raise the price → the data shows that customers are willing to pay this much, there is no point in selling it cheaper.
Product 3: a more significant increase → you probably underpriced it so far.
The model is not a “discount machine”, and it does not always decrease – it moves the price to where you make more money while the customer remains satisfied.
The question is fair:
“Wouldn’t it be better to work with fixed prices? Won’t this disturb customers?”
As humans, it is impossible to determine where the optimal price is.
Even with a few products it’s not transparent which price would generate more profit.
With hundreds or thousands of products it becomes completely impossible.
Let’s look at a single product.
In the following table you can see the sales data:
What do you think?
Based on this, can you say for sure that in the next period you should:
raise the price or lower it?
if yes, by exactly how much?
and how would you justify this, not just based on a gut feeling?
As humans, at this point we can only guess.
A fixed price is comfortable, but not optimal.
It doesn’t react to changes in demand, seasonality, campaigns, inflation.
The market is constantly moving.
Christmas, Black November, Black Friday, Easter, trends, competition, changes in logistics. It is impossible to follow all of this continuously and correctly by hand.
Discounting based on human decisions is almost unmanageable.
You don’t know:
when to discount,
by how much to discount,
for how long to discount,
and how to avoid damaging your brand’s credibility.
The model solves these problems:
it works with a win–win mindset:
higher, more stable profit for you,
understandable, fair prices for the customer,
and it does not go into price-cutting, does not produce unnecessary “-70%” shocks.
Important: the model does NOT rely on price-cutting, it does not pull the market down, it makes it more orderly.
Before developing our own model, we analyzed the pricing solutions available on the market for several months (rule-based systems, competitor monitoring, A/B testing, AI/ML pricing). Our conclusion was clear: they are not suitable for providing stable, provably good prices in the long term.
Rule-based and A/B testing systems
you have to come up with the rules manually → the same burden, just on a more complicated interface,
they do not learn from real customer behavior,
if the market changes, they can easily pull in the wrong direction,
every webshop is different, yet the same template is forced on all.
Result: lots of settings, little real optimization.
Competitor-monitoring systems
start a price war spiral: everyone goes lower and lower, margins disappear,
ignore your own data (your brand, your customers’ purchasing power),
copy the bad decisions of others instead of looking for an optimum,
in the long term they lead to price erosion and loss of trust.
Result: you don’t win, you just lose more slowly.
Classic “AI” / ML pricers
too many parameters, too little data → overfitting, they memorize the past, but are uncertain about the future,
they are sensitive to incomplete, noisy data – and sales data is very noisy,
they give unstable, jumping prices, or simply make random decisions.
Result: they may look impressive, but they do not provide reliable, transparent decisions.
If only you use the model:
You don’t get dragged into the price-cutting spiral. The model does not follow competitors’ prices, it calculates the current optimum from your own customers’ behavior.
Result: higher, more stable profit, fewer unnecessary discounts – without undercutting.
If your competitors also start using it:
This is even better for you. Due to the logic of Nash equilibrium, the system does not underbid each other, but looks for everyone’s own profit maximum. Outwardly, the market moves out of the price-cutting spiral and upwards into the optimal price band – where profit is sustainable and demand is stable.
It’s not collusion: there is no data sharing, we calculate separately and independently for everyone; the equilibrium is a natural market outcome.
“Will prices shoot up?” No.
Prices are limited by the upper bound of demand. Today’s market, because of price undercutting, is typically operating below the optimum; the model pulls this back into the unknown but existing optimal band.
In summary:
You win even alone – and the more people use it, the faster price wars die out and the market converges to the truly profitable equilibrium prices, without coordination.
Three things are needed:
Unit cost (gross)
Current selling price (gross)
Sales data from the last 12 months:
when the sale took place,
at what price it was sold,
the quantity sold.
A simple example:
The model looks at the prices at which the product has been sold so far and at what intervals.
It determines the next step that moves the price closer to the optimum.
It repeats this step hourly, daily, or at the chosen interval until the price settles into a stable, good range.
If there is no unit cost → we do not optimize that product.
There is a blacklist: any fixed-price product, manufacturer-specified price, etc. can go here.
The unit cost acts as a minimum price: under no circumstances will the model go below it.
And because the goal is profit, the model will generally not “crawl down” close to the unit cost either, since that would be a “bad deal” as well.
While running the system, we handle your data securely:
we do not store the data,
once the new prices are written back into your store, the runtime environment and the data are deleted.
If you want extra security, you can transform your product IDs (SKU, item number) in advance with one-way encoding (hash), and only send them to us in this “unreadable” form.
This means that:
we do not see which specific product in your store “Product 1” really is,
and due to one-way encoding, no one can reverse the identifiers back to the original products.
Example:
let’s say the store name is “Store 1”, and one of the products is called “Product 1”.
instead of using “Product 1” as the product identifier in the table,
your store transforms the identifier using a one-way encoding function (hash) and only sends it to us like this: “146532e9958445fd666928434c1bcb4c0884f56cd53cbb2c0a6d405534da3522”
So we only see that there is an item called “146532e9958445fd666928434c1bcb4c0884f56cd53cbb2c0a6d405534da3522”, but we cannot reverse that this is “Product 1” in “Store 1”.
Amazon updates roughly every 7 minutes – this shows how important dynamics are.
In most cases we recommend hourly optimization.
There is the possibility to run:
with any hourly step (1 hour, 2 hours, 4 hours, 13 hours, 18 hours, etc.),
daily,
weekly – adapted to your system’s refresh capabilities.
This is not a problem. Advertising platforms (Google Merchant Center, Facebook Meta, etc.) can automatically update prices as soon as they change in the store. If frequent updating is not possible, then the optimization must be set to the lowest refresh rate of the ad platform.
Example:
if Google Merchant Center updates prices every 6 hours,
then we run the model every 6 hours,
so the price shown in the ad and in the webshop will match.
Especially important:
the model does not replace your ads; it strengthens them,
your advertising campaigns should continue to run while the model is in use.
No. And this is very important. The system does not rewrite all prices in every cycle.
If for a product it has already found the optimal price and it continues to perform well, it leaves it unchanged – even through many runs.
It only intervenes if there is a real reason based on the data:
sales are slowing down,
demand or season changes,
it becomes visible that a slightly higher or lower price would bring more profit,
a controlled, short “promotional” price level is needed to stimulate demand.
So: it is stable where the price is good – it changes intelligently where you would be leaving money on the table.
The model often does not give “nice round” prices, but for example:
14.76
1327
2593 etc.
We have built in:
psychological rounding ending in 9 (e.g. 1439, 1.99),
rounding ending in 0 or 5 (e.g. 1405, 120),
but we only recommend using these in special cases.
Based on our research:
in online shopping, the ending of the price does not meaningfully influence the decision,
what matters more is that the price is logical, consistent and competitive.
Psychological price-ending strategies may still have some relevance in offline (physical) stores, though we have not found any research that convincingly confirms this for us. If you want, the model can adapt to the desired appearance – but from a profit perspective, the “nice number” is secondary.
There are products that are new, or simply have not yet had any sales. As humans, we usually:
“roughly” guess a price,
and then wait – for weeks or months – without really learning anything from it.
By contrast, the model explores deliberately:
it calculates a price interval (range) that is realistically possible for this product,
and within this range it randomly chooses prices over successive runs.
This is not a bug; from an information-theoretic point of view, it is the best strategy:
this way we can most efficiently map out what price customers accept.
Example:
The starting price of the product: 150 USD, no sales yet.
At the first optimization, the model suggests 180 USD.
At first glance you might say: “This is too high, no one will buy it at this price.”
But in reality, we don’t know: they might buy it at this price too. We need to explore this.
If there is still no sale, at the next run the model might suggest 126 USD.
This is a lower price; we then see whether demand starts at this level.
The point is:
as humans, we do not know where the price is that customers accept,
so at the beginning the model deliberately tries both higher and lower directions,
as soon as at least two sales happen, there is already enough information for the system to:
leave the “exploratory” phase,
and precisely, stably optimize the product’s price.
Thus for new products we are not shooting in the dark, but quickly and in a controlled way find the price level where it brings the most profit – without losing months to a bad starting price.
Discounting is impossible to manage manually:
we don’t know the optimal discount price,
we don’t know when we should discount,
we don’t know how long we should discount,
and the discounting style typical on the market is often harmful:
Original price: 2500 → Discount price: 1500: the customer feels the product was overpriced before.
Everything is “red” and discounted: this signals that the store is in trouble.
We discount the wrong product at the wrong time and by the wrong amount → lost profit.
Our model works differently:
it does not trick by using a separate “discount price” table,
but modifies the product’s price so that it:
drops just enough to boost sales,
stays low just as long as needed,
and does not damage the brand or create distrust.
The model:
knows when it’s worth changing the price,
knows what size of adjustment is justified,
and keeps the “discount” level exactly as long as it generates profit.
Takes the burden off your shoulders → controlled, data-driven discounting.
For us, the unit cost is:
the minimum threshold, we do not go below this,
our starting point, our base.
We do not recommend that you “pour” all your store’s overheads, wages, utilities, shipping costs into the unit cost.
Example:
The unit cost of Product 1: 1340
You decide that with all costs included it should be 1540.
But it may be that the product’s optimal selling price would be 1450.
If you give 1540 as the purchase price, the model cannot go down to 1450,
because it thinks that would be a loss.
By doing this you actually:
block the good solution, and
signal: “this product was purchased expensively or designed with a bad cost structure”.
The model:
does not perform magic,
cannot fix a bad purchase price or bad cost structure,
but it does help you get the maximum profit out of the existing situation.
Your store’s operating costs cannot directly dictate individual product prices.
The model cannot generate infinite profit; therefore it will not be able to eliminate the consequences of bad cost planning and the low margins of products with non-optimal purchase prices.
Planning fixed costs is the owner’s task; we put your pricing in order.
We don’t turn it on for your entire store right away. We work in three steps:
Level 1 – technical test
the model selects a few (typically 2–20) of your fastest-moving products,
for one day we only optimize these,
and together with you we check the pricing of these products.
Level 2 – measurement
here it is necessary to have the unit cost of all products entered into your system,
we expand the circle (typically 5–50 strong products),
the model selects the control group of products,
we carry out the measurement, which can last from a few weeks up to a few months depending on product turnover speed.
Level 3 – full optimization
we enable optimization for the entire product range.
Thus:
we prove ourselves in a small, risk-free setting,
we transparently show what the model does,
and we only move forward if the result is clearly visible.
Full product range optimization starts within a few days, but at most within a week.
The model runs on stable infrastructure (Google Cloud), it is scalable and reliable.
It always strives to move prices toward a better optimum.
It follows micro and macro trends, seasonality, and continuously adapts.
It is a system developed specifically for dynamic pricing, not a by-product.
We measure its impact:
we produce automatic reports,
we show how much extra profit comes from the model,
we use product scoring to indicate which products are pulling the store and which are harming it.
Human pricing is no longer enough today. Too much data, too fast a market, too much emotion.
The HONAMA dynamic pricing model replaces this with a system that:
is measurably better in terms of results,
is fair to the customer,
is transparent for the owner,
and does not promise miracles – only the level of precision that is unattainable by hand.
Correctly measuring a store’s performance is an extremely complex task for humans. Most owners work instinctively, with simple comparisons – “was revenue higher than last month?” – while they do not take into account market effects: seasonal fluctuations, inflation, consumer sentiment, external events.
Without these aspects, measurement almost always gives a distorted picture: it may look like the store is performing better or worse while only the environment has changed.
In our measurements, we treat the store’s own performance and market effects together, and we use professional, statistically grounded methods that filter out randomness and external noise. Using hypothesis tests based on complex probability distributions is critically important to see through the informational noise and to determine whether a growth is due to randomness or actually the effect of optimization.
The result of this: real, demonstrable performance indicators suitable for business decisions.
Net profit / day (calculated from orders)
Profit stability (reduction of volatility)
Relative performance compared to the market (if the market falls, how much less do you fall?)
We use two mutually reinforcing methods:
Before–after (temporal) measurement with permutation hypothesis testing
we compare the profit of time periods of equal length before and after optimization,
the permutation test tells us what the probability is that the difference is due to chance alone.
Parallel control vs. optimized group
during the same time period we compare optimized products to control (non-optimized) products,
so market effects cancel out completely (same time, same environment).
Do not stop advertising in the period before optimization and during the measurement window.
Try not to change logistics, discounting policy, category offering, or cost calculation within the measurement window.
Ideally: at least 1 month of stable baseline before starting.
The point: we do not measure on the basis of “is revenue higher than last time”, but we show how much the model brings in a market-corrected, statistically proven way – even in a negative market. For a store, the greatest value is knowing exactly what works and what doesn’t – this is the certainty we provide.
It is important to emphasize that the model does not only work well in a growing market, but also when the market is falling.
Most simple measurements fail here: if traffic or profit decreases, they automatically interpret it as a failure, while the market may be performing much worse.
Example:
if the market as a whole falls by 20%,
but with the model your store falls only by 10%,
then this is not “just” a smaller loss, but a performance 10 percentage points better than the market.
In other words, in an environment where others lose 20%, you effectively gain compared to them with a 10% drop.
Biocity Hungary Kft. (Szeged, bio products, 3200+ active products):
after the start of dynamic pricing, over an 8-day period
+52% net profit growth compared to the previous 8 days,
over nearly 3 months, 2.8x profit compared to the previous period of equal length,
which is roughly 62% average monthly profit growth,
and according to statistical analysis, the effect has more than a 95% probability of not being due to chance.
This is not a theoretical promise, but live data.
