Yep, this can lead to troubling bugs (or exploits) when attributes can change as well -- e.g. if you happen to level-up while a buf is enabled.

So:
3. Represent the stats as a stack of operations.
e.g. a stack for your speed stat might contain:
*Slow: Subtract 10
*Base: Set 50
When you evaluate that from the bottom up, you get (50-10) = 40.

When you cast the slow spell, the new instance of that spell adds an operator to speed's stack and retains a handle to it. When the spell ends, it uses that handle to remove the operator that it added earlier.

If the player levels up, you can modify the 'base' item in the stack above, and it will automatically allow you to recompute the new resulting speed value, including buffs.

You can keep a cache of the speed attribute as well as this stack, and update the cache whenever the stack is modified... But never change that cached value -- always change the stack of operators, and let the stack update the cache with the new value.

I all solutions above are great solutions for an RPG. Which one is better may depend on the requirements of your game.

However, one thing they all have in common is that the base speed is stored with the character and not the modified speed.

The modified speed is calculated when the speed is requested

A simple example of a GetSpeed function could look like the following:

class character {
    private int _baseSpeed;
    private bool _isSlowed;
    
    public int GetSpeed() {
        int speed = _baseSpeed;
        
        if(_isSlowed) {
            speed -= 10;
        }

        //return 0 if calculated speed < 0, otherwise return calculated speed
        return speed < 0 ? 0 : speed;
    }
}

Hope this adds some clarification, if it was not already clear.

Good luck on your RPG!

Massively simplify the amount of state you're modifying by using more code and data.

Store a base/"true" stat. Do not change this, unless some effect is supposed to _permanently_ alter the base state.

Keep a running/"current" stat value modified by effects. This should ideally be a cache, e.g. you should be able to recalculate this from all current effects. You can either invalidate the cache whenever an input changes (e.g., an effect is added/removed, the base stat changes, etc.) or design the system to require few invalidations (if effects are expensive to calculate). If doing the latter, keep this stat simple - do _not_ clamp the value here (let it go negative!).

Provide a public access for the stat that hides the differences between base, current, and clamped values. This accessor is what all your public interfaces should use, with a few exceptions (UI probably has places that it wants to show the base stat or compare and contrast the current with the base, so make a public readonly accessor for that too for UI).

Now you don't have a source of bugs. Base is base and doesn't get changed much (no source of bugs), current is recalculated when needed or has composable math (no source of bugs), and the accessor does any last-minute logic to ensure the value matches game rules (no source of bugs).

3. Represent the stats as a stack of operations.
e.g. a stack for your speed stat might contain:
*Slow: Subtract 10
*Base: Set 50
When you evaluate that from the bottom up, you get (50-10) = 40.

A similar approach, and the one we took, was to have two classes for characters (or enemies, but lets focus on characters). One is called GlobalCharacter, and this is the permanent class object that maintains all of the character's stats. When we enter a battle, we create a BattleCharacter class which contains a copy of all the GlobalCharacter stats. We don't touch the global character stats in battle, so we always have the original values. When the battle ends, we set the stats on the GlobalCharacter to those of the BattleCharacter to reflect changes like HP or mana.

This also has the convenience of allowing us to restore the conditions when a battle began, which is something we take advantage of as we allow the player to restart a battle a number of times if they are defeated.

A good design pattern for this is the Decorator pattern. Applied to this problem, you can think of it as an object-based formalization of Hodgeman's stack suggestion, or some of the other suggestions here.

How I use it is this --

First, in my case, characters' stats are fixed per-level, and stored in a read-only array-of-structs -- there's one struct for each character and for each level. The character data structure points to this for its base stats. Weapon, armor, and item stats are similar, but separate -- You can think of all of this as being a big dictionary, or set of spreadsheets or database tables -- in fact, that's usually how I author this data, which gets fed into the build (or could be read from a file a load-time).

*side note* If you have non-static leveling (e.g. where the user allocates their own points when they level up) you would manage this base data differently, but the basic pattern and application stays the same. You have a couple obvious choices: One is that you do away with the dictionary and just have a base-stat data structure for each character/item that you modify directly; The second way is that you still have the dictionary of read-only base stats, and you implement the user-allocated points themselves using this pattern (just apply them before any other buffs/debuffs/effects).

Now you apply the decorator pattern to that base stat structure -- basically, you have a decorator that presents the same interface as the structure itself, and when a decorator is applied, the character points to the decorator, and the decorator points to the base stat. This means the character reads from the decorator, which is where it gets its opportunity to modify the base values. What about multiple effects/buffs/debuffs? Simple! Since the decorator has the same interface as the base state, a decorator can point to either the base state or another decorator as long as they derive from the same interface and you use a pointer to the base class -- you can stack them arbitrarily deep, and each decorator will modify the sum of the base stat and all the decorators that came before it.

This might sound like an awful lot of engineering for something so simple, but its very robust and the pattern isn't terribly complicated. One of the things that makes it nice is that it really decouples the implementation of effects from one another and from the base stats structure. They don't need to know anything about what the other states look like, how many there are, or what they can do. It communicates only the language of the character stats (though, in my implementation, I have a side-channel that allows an effect to know if a certain effect has already been applied -- for items, think of a set of armor that's stronger when all its pieces are equipped.)

Also worth noting, is that the reason I like to pass in the entire character stat-set, rather than having individual decorators for each stat is that it makes it much easier to have multi-dimensional effects that modify several stats, or which take into account multiple stats in determining its value (e.g. a character's defense against spells might be modified based on both their magic resistance and their intelligence) -- its also a whole lot less to manage and a whole lot less code to write.

To talk about implementation of the interface for a bit, it'll depend on whether you're programming in a language with properties (like C#) or without (like C++) -- if you have properties, then you can use them to implement getters for each base stat (and setters, if you don't go the dictionary of read-only structs route). Otherwise, you need to implement member functions to get the values out.

Another good shortcut to know about is that in C++ a virtual function can have a default implementation -- so what I do is when I define the base class for the decorators is give the virtual getter a default implementation that just reads the value form whatever it points to and passes it along unmodified. That way, when an implementation of a particular decorator only modified one stat, I only have to implement that one stat's getter without any other boilerplate. Not only does it saves me typing, it removes the opportunity to make mistakes.